A multi-parameter fused electric actuator health degree evaluation method and system

Abstract

The application discloses a kind of multi-parameter fusion's electric actuator health degree evaluation method and system, belong to equipment state monitoring and management technical field, including: acquisition electric actuator multi-source parameter, valve load working condition is encoded as working condition fingerprint vector, obtain multi-parameter time series dataset;Extract torque, displacement, current three parameters, combined with valve type coding constructs three-coupled phase space and reconstruction, extract degenerate manifold feature vector, with parameter as node, causal direction as directed edge, causal strength as edge weight, combined with working condition fingerprint vector constructs degenerate causal directed graph, calculates topological degeneration index vector;With degenerate physical mechanism as priori, manifold feature and topological index as observation evidence, by Bayes inference obtains the dynamic weight vector of each parameter health degree contribution, and combined with the cumulative action number in working condition fingerprint calculates comprehensive health degree, determines degradation grade and generates maintenance strategy;The application improves the accuracy of electric actuator health degree evaluation.

Description

Technical Field

[0001] This invention relates to the field of equipment condition monitoring and management technology, specifically to a method and system for assessing the health of electric actuators using multi-parameter fusion. Background Technology

[0002] Electric actuators are widely used in industrial pipeline systems such as petrochemical, power, and metallurgy to drive valves and regulate their opening. Their operating status directly affects the safety and regulation accuracy of the pipeline system. Most existing methods for assessing the health of electric actuators have the following shortcomings.

[0003] Existing methods typically rely on a single type of parameter for threshold judgment, failing to comprehensively utilize the coupling relationship between electrical, mechanical, and state parameters. This results in insufficient sensitivity to early, minor degradation, easily leading to missed detections. Multi-parameter fusion methods often use fixed or empirical weights to weight and sum the features of each parameter, failing to consider the dynamic changes in the contribution of different parameters to health as the degradation stage and operating conditions change. This leads to significant deviations in assessment results when operating conditions change or degradation modes change. Furthermore, the lack of characterization of causal propagation relationships between parameters makes it difficult to distinguish between the source of degradation and the intermediate links in its propagation, failing to provide information on the root cause and propagation path of degradation for maintenance decisions, resulting in insufficient targeting of maintenance strategies. Finally, the differences in valve load conditions are not adequately considered. When the same assessment model is applied to different valve types, the fundamental differences in the time delay characteristics of torque-displacement coupling mean that using uniform feature extraction parameters will cause distortion in dynamic feature extraction, thus affecting the accuracy of health assessment.

[0004] Based on this, the present invention designs a multi-parameter fusion method and system for assessing the health of electric actuators to solve the above problems. Summary of the Invention

[0005] In view of the above-mentioned shortcomings of the existing technology, the present invention provides a method and system for health assessment of electric actuators based on multi-parameter fusion.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A multi-parameter fusion method for assessing the health of electric actuators includes the following steps: Step S1: Collect multi-source parameters of the electric actuator, and at the same time obtain the valve load condition context information and encode it into a condition fingerprint vector. Perform time-series alignment on the multi-source parameters to obtain a multi-parameter time-series dataset. Step S2: Extract three parameters, torque, displacement and current, from the multi-parameter time series dataset. Combine the valve type encoding in the working condition fingerprint vector to construct a three-element coupled phase space and reconstruct the phase space. Extract the degenerate manifold feature vector of the electric actuator. Step S3: Based on the causal direction between parameters in the multi-parameter time series dataset, and using the working condition fingerprint vector as a condition variable to exclude false causality, construct a degenerate causal directed graph with each parameter as a node, causal relationship as a directed edge, and causal strength as the edge weight. Step S4: Perform topological analysis on the degenerate causal directed graph and calculate the topological degradation index vector representing the characteristics of degradation propagation; Step S5: Using the degradation physical mechanism of the electric actuator as a priori constraint, and the degradation manifold feature vector and topological degradation index vector as observational evidence, determine the dynamic weight vector of each parameter's contribution to health through Bayesian inference. Step S6: Based on the dynamic weight vector, the degenerate manifold feature vector and the topology degradation index vector are weighted and fused, and combined with the cumulative number of actions in the working condition fingerprint vector, the comprehensive health of the electric actuator is calculated. The degradation level is determined according to the comprehensive health and the corresponding maintenance strategy is generated.

[0007] Furthermore, the multi-source parameters include electrical parameters, mechanical parameters, and state parameters. Electrical parameters include drive current, terminal voltage, and power factor; mechanical parameters include output shaft torque, valve stem displacement, and angular velocity; and state parameters include motor winding temperature, reducer oil temperature, and housing vibration acceleration. Valve load condition context information includes medium pipeline pressure, ambient temperature, valve type code, opening command value, and cumulative number of actions; The entire stroke from receiving the opening command to reaching the target opening degree of the electric actuator is taken as an action cycle. Within each action cycle, the sampling points of each parameter are synchronized and aligned with a unified time reference. Each action cycle in the multi-parameter time series dataset is associated with the corresponding working condition fingerprint vector.

[0008] Furthermore, a ternary coupled state vector is constructed using the output shaft torque, valve stem displacement, and drive current in the multi-parameter time series dataset as three original components; The optimal delay time between each component in the ternary coupled state vector is determined by the mutual information method, and the optimal embedding dimension of the phase space is determined by the Cao method. The phase space of the ternary coupled state vector is reconstructed using the optimal delay time and the optimal embedding dimension to obtain the sequence of trajectory points in the reconstructed phase space. Based on the valve type code in the working condition fingerprint vector, the torque-displacement coupling time delay characteristics corresponding to the current valve type are determined, and the search range of the optimal delay time is adaptively adjusted accordingly.

[0009] Furthermore, on the trajectory point sequence, the maximum Lyapunov exponent, the correlation dimension of the reconstructed phase space, and the volume shrinkage rate of the reconstructed phase space are calculated, and the three are concatenated to obtain the degenerate manifold feature vector.

[0010] Furthermore, for each parameter time series in the multi-parameter time series dataset, detrending and variance normalization are performed using the action cycle as the segment unit; For any two normalized parameter time series, the conditional Granger causality test is used to determine the causal direction relationship between them. The conditional Granger causality test uses the medium pipeline pressure, opening command value and ambient temperature in the operating condition fingerprint vector as condition variables to eliminate false causal associations caused by changes in operating conditions. When the causal relationship between two parameters passes the significance test, a corresponding directed edge is established in the degenerate causal directed graph, and the direction of the directed edge is the causal action direction; the F-statistic obtained from the conditional Granger causality test is normalized and used as the edge weight of the directed edge. The health baseline values ​​are reference values ​​for the maximum Lyapunov index, correlation dimension, and volume shrinkage rate calibrated under healthy conditions during the initial operation of the electric actuator. When the maximum Lyapunov exponent in the eigenvector of a degenerate manifold exceeds its corresponding health baseline, the decision threshold for significance testing is lowered to enhance the ability to capture weak causal relationships.

[0011] Furthermore, in the degenerate causal directed graph, the node with zero in-degree and the largest out-degree is identified as the degenerate source node, which represents the parameter channel where the degeneracy first occurs; Starting from the degenerate source node, calculate the shortest directed path length from it to all end nodes with an out-degree of zero, and take the average of all shortest directed path lengths as the propagation path length. The proportion of downstream reachable nodes of the degenerate source node to the total number of nodes in the degenerate causal directed graph is calculated as the branch diffusion degree. The number of directed loops in a degenerate causal directed graph is detected, and the normalized number of directed loops is used as the loop risk index. The identifier of the degradation source node, the propagation path length, the branch diffusion degree, and the loop risk index are combined into a topology degradation index vector.

[0012] Furthermore, a priori order relations are established based on the degradation physical mechanism of the electric actuator: abnormal temperature rise of the motor winding precedes output shaft torque fluctuation, output shaft torque fluctuation precedes valve stem displacement deviation, and valve stem displacement deviation precedes abnormal drive current. The priori order relations are encoded as prior distribution parameters of the dynamic weight vector. The expected value of the prior distribution parameters decreases monotonically from front to back along the priori order relations, so that the initial weight corresponding to the parameter ranked earlier in the priori order relations is greater than the initial weight corresponding to the parameter ranked later. Using the degenerate manifold feature vector and topological degradation index vector as current observation evidence, Bayesian update is performed on the prior distribution parameters to obtain the posterior distribution. The expected value is taken from the posterior distribution as the final weight of each parameter in the dynamic weight vector. When the loop risk index in the topology degradation index vector is greater than zero, a penalty decay factor is applied to the weights corresponding to each parameter involved in the directed loop to suppress the weight concentration caused by the positive feedback loop.

[0013] Furthermore, based on the degree of deviation between each topological invariant in the eigenvector of the degenerate manifold and the health baseline, the health sub-score of the degenerate manifold is calculated; Based on the degree of deviation between the propagation path length, branch diffusion degree, and loop risk index in the topology degradation index vector and the topology benchmark value corresponding to the health benchmark value, the causal topology health sub-score is calculated. The overall health score is obtained by weighting and summing the health scores of the degenerate manifold and the causal topological health sub-scores using a dynamic weight vector. The overall health score is corrected based on the cumulative number of actions in the working condition fingerprint vector: when the cumulative number of actions exceeds the preset life cycle ratio, a decreasing correction coefficient is applied to the overall health score. Based on the revised comprehensive health value range, electric actuators are classified into five degradation levels: healthy, slightly degraded, moderately degraded, severely degraded, and at risk of failure, and corresponding maintenance strategies are generated.

[0014] Furthermore, using the numerical sequence of comprehensive health score over multiple consecutive action cycles as input, a grey prediction model is employed to extrapolate and predict the future evolution trend of comprehensive health score. Based on the prediction results, the number of remaining operating cycles required for the overall health status to drop below the threshold corresponding to the failure risk level for the first time is calculated as the estimated remaining usable life of the electric actuator. When the estimated remaining usable life is less than the preset safety margin period, the degradation level will be increased by one level and the maintenance strategy will be upgraded. The remaining usable life estimate is correlated with the cumulative number of actions in the operating condition fingerprint vector to generate the full life cycle degradation curve of the electric actuator.

[0015] A multi-parameter fusion-based health assessment system for electric actuators, the system comprising: The multi-parameter acquisition module is used to acquire multi-source parameters of the electric actuator, obtain valve load condition context information and encode it into a condition fingerprint vector, and perform time-series alignment of the multi-source parameters to obtain a multi-parameter time-series dataset. The phase space reconstruction module is used to extract three parameters—torque, displacement, and current—from the multi-parameter time series dataset. Combined with the valve type encoding in the working condition fingerprint vector, it constructs a three-dimensional coupled phase space and performs phase space reconstruction to extract the degenerate manifold feature vector of the electric actuator. The causal graph construction module is used to construct a degenerate causal directed graph based on the causal direction between parameters in a multi-parameter time series dataset, and to exclude false causality by using the working condition fingerprint vector as a condition variable. The graph has each parameter as a node, causal relationship as a directed edge, and causal strength as the edge weight. The degradation feature analysis module is used to perform topological structure analysis on the degradation causal directed graph and calculate the topological degradation index vector representing the degradation propagation characteristics. The contribution weight analysis module is used to determine the dynamic weight vector of each parameter's contribution to health by using the degradation physical mechanism of the electric actuator as a prior constraint, the degradation manifold feature vector and the topological degradation index vector as observational evidence, and Bayesian inference. The maintenance decision generation module is used to perform weighted fusion of degraded manifold feature vector and topological degradation index vector based on dynamic weight vector, and combine it with the cumulative number of actions in the working condition fingerprint vector to calculate the comprehensive health of electric actuators. Based on the comprehensive health, the degradation level is determined and the corresponding maintenance strategy is generated.

[0016] Compared with the prior art, the beneficial effects of this invention are as follows: By constructing a three-dimensional coupled phase space based on the output shaft torque, valve stem displacement, and drive current, and performing nonlinear dynamic analysis, it is possible to capture progressive degradation modes such as increased internal friction and widened transmission clearance, which are difficult to reflect by existing single-parameter statistical feature methods. The sensitivity of identifying early slight degradation is significantly better than existing methods based on statistical thresholds.

[0017] The constructed causal directed graph of degradation can identify degradation source nodes and propagation paths. Compared with existing methods that only output health values ​​without the ability to locate the root cause of degradation, it can provide maintenance personnel with more targeted maintenance strategies.

[0018] We employ Bayesian dynamic weighted fusion under the prior constraints of physical mechanisms. The weights are adaptively updated with the observed evidence and causal topology. Compared with the existing fixed-weighted fusion method, the stability and accuracy of the evaluation results are significantly improved in scenarios such as working condition switching and degradation mode changes.

[0019] By combining the grey prediction model to extrapolate the evolution trend of overall health and linking the prediction results with the cumulative number of actions, a full life cycle degradation curve that can be used for horizontal comparison of equipment of the same batch and model is formed, providing trend and group analysis capabilities that are not available in existing single-point assessment methods for equipment ledger management and predictive maintenance. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0021] Figure 1 This is a flowchart of the multi-parameter fusion-based health assessment method for electric actuators according to the present invention; Figure 2 This is an architecture diagram of the multi-parameter fusion electric actuator health assessment system of the present invention. Detailed Implementation

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

[0023] Example 1: Figure 1 The present invention provides a multi-parameter fusion-based health assessment method for electric actuators, comprising the following steps: Step S1: Collect multi-source parameters of the electric actuator, and at the same time obtain the valve load condition context information and encode it into a condition fingerprint vector. Perform time-series alignment on the multi-source parameters to obtain a multi-parameter time-series dataset.

[0024] Step S1 includes the following steps: During the operation of the electric actuator, multi-source parameters are collected, and valve load condition context information is acquired simultaneously. The multi-source parameters include three categories: electrical parameters, mechanical parameters, and status parameters. Electrical parameters include drive current, terminal voltage, and power factor, reflecting the electromagnetic load state of the motor; mechanical parameters include output shaft torque, valve stem displacement, and angular velocity, reflecting the mechanical transmission state of the actuator; status parameters include motor winding temperature, reducer oil temperature, and housing vibration acceleration, reflecting the thermal and structural state of the actuator. Valve load condition context information includes medium pipeline pressure, ambient temperature, valve type code, opening command value, and cumulative number of actions. The valve type code is used to distinguish between valves with different structural forms, such as gate valves, ball valves, and butterfly valves.

[0025] Taking the complete stroke of the electric actuator from receiving the opening command to reaching the target opening as one action cycle, record the total number of data collected during this action cycle. There are sampling points, and the sampling time interval is . Within each action cycle, the sampling points of each parameter are synchronized and aligned using a unified time reference, meaning all parameters are sampled at the same time. ( Each sensor has a corresponding sample value, eliminating timing misalignment caused by acquisition delays from different sensors, and obtaining a multi-parameter timing dataset corresponding to the action cycle.

[0026] The valve load condition context information within this action cycle is encoded into a condition fingerprint vector: ; in, For operating condition fingerprint vectors, This represents the average value of the pressure in the medium pipeline. The average value of the ambient temperature is the sampled value. Valve type coding, This is the opening command value. This represents the cumulative number of actions.

[0027] Valve type coding assigns category identifiers to different valve structural forms. This refers to the cumulative number of actions, i.e., the total number of action cycles completed by the electric actuator since it was put into operation. Operating condition fingerprint vector. Each parameter is associated with a time-series dataset of the action cycle, serving as the basis for subsequent steps such as distinguishing work scenarios, eliminating false causes, and correcting health status.

[0028] Step S2: Extract three parameters, torque, displacement and current, from the multi-parameter time series dataset. Combine the valve type encoding in the working condition fingerprint vector to construct a three-dimensional coupled phase space and reconstruct the phase space. Extract the degenerate manifold feature vector of the electric actuator.

[0029] Step S2 includes the following steps: Extracting output shaft torque from a multi-parameter time-series dataset Valve stem displacement and drive current Three parameters ( The three parameters are chosen as the original components of the three-dimensional coupled state vector because torque, displacement and current are coupled from the three dimensions of mechanical load, geometric stroke and electromagnetic response to reflect the action process of the actuator. The distortion of their coupling relationship can show degradation characteristics such as increased mechanical clearance and abnormal transmission resistance earlier.

[0030] The mutual information method is used to determine the optimal delay time between each pair of the three components. For any component sequence Define itself in delay Average mutual information under: ; in, For delay The average mutual information value under the following conditions For component sequences At any moment The probability distribution of the values, For component sequences At any moment The probability distribution of the values, For a moment Values ​​and Time The joint probability distribution of the component sequences of the values.

[0031] Pick Follow The first local minimum that appears after the increase corresponds to As candidate optimal delay times for this component, the candidate optimal delay times for the torque, displacement, and current components are denoted as follows: , , .

[0032] Before determining the candidate optimal delay time, a mutual information search interval must first be given. In existing technologies, the search interval is typically determined empirically by taking a fixed multiple of the number of sampling points. This invention adaptively adjusts the search interval. Specifically, it first calculates the average rate of change of torque relative to displacement within the current action cycle: ; in, This represents the average rate of change of torque relative to displacement within the current action cycle. for The output shaft torque sample value at the sampling time. for The output shaft torque sample value at the sampling time. The sampling time interval, This represents the number of sampling points in this action cycle.

[0033] It reflects the degree of torque change over time within this action cycle. The larger the value, the faster the torque changes, and the smaller the value, the smoother the torque changes.

[0034] Furthermore, the valve type encoding in the current operating condition fingerprint vector is further refined. Filter out those related to the historical action cycle database For the same historical operating cycle, update the torque change rate reference benchmark for this valve type incrementally: ; in, To accumulate to the number Reference benchmark for torque change rate after one cycle of the same type of action. To accumulate to the number Reference benchmark for the same type of action cycle For the first The rate of change of torque calculated from a cycle of the same type of action. This refers to the sequence number of the cumulative action cycles for this valve type in the statistics.

[0035] The reference benchmark is continuously updated as data accumulates, reflecting the torque variation characteristics of this valve type in actual operating scenarios.

[0036] Based on this reference benchmark and the preset initial search range The initial interval can be set according to a conventional empirical ratio of the number of sampling points, serving as a general starting point for feature extraction. The adaptive scaling factor is then calculated. ; in, For adaptive scaling factor, To accumulate to the number Reference benchmark for torque change rate after one cycle of the same type of action. This represents the average rate of change of torque relative to displacement within the current action cycle.

[0037] When the valve type code corresponds to a gate valve, its stroke is relatively large and the torque change is more gradual, according to actual measurements. Typically smaller than the historical reference benchmark of this type, therefore Press at this time Increase the upper limit of the search range to accommodate the gradual torque change characteristics of the gate valve during its large stroke; when the valve type code corresponds to a ball valve, its torque changes rapidly within a small stroke range, as measured in actual tests. It is usually greater than the historical reference benchmark, therefore The upper limit of the search interval is correspondingly reduced to accommodate the rapid torque mutation characteristics of ball valves. The scaling factor is calculated based on measured data and a self-updated historical statistical benchmark, enabling adaptive adjustment of the search interval for different valve types. Within the adjusted search interval... Recalculate each component internally. The curve is then analyzed, and local minima are found to obtain the final optimal delay time. , , .

[0038] The Cao method is used to determine the optimal embedding dimension of the phase space. For a given embedding dimension... ,definition: ; in, For embedding dimension The following is constructed from the ternary coupled state vector. One reconstructed vector, for The nearest neighbor vector in the reconstructed phase space For Euclidean distance, This represents the representative delay time between components. Embedding dimension Lower average distance ratio Embedding dimension Lower average distance ratio To describe With dimension Increased relative rate of change This represents the number of sampling points in this action cycle.

[0039] when Follow When the value increases and tends to stabilize, take the minimum value that tends to stabilize. Incrementing the value by one as the optimal embedding dimension .

[0040] With optimal delay time and optimal embedding dimension Phase space reconstruction of the ternary coupled state vector yields a sequence of trajectory points: ; Three degenerate manifold features are computed on this trajectory point sequence. The maximum Lyapunov exponent is estimated using a small-data-volume algorithm: ; in, The maximum Lyapunov exponent. For trajectory points The distance between it and its nearest neighbor trajectory point For trajectory points The distance between it and its nearest neighbor trajectory point The sampling time interval, This represents the number of trajectory points included in the statistics.

[0041] The larger the value, the faster the separation speed of adjacent trajectories, the higher the dynamic complexity and uncertainty of the system, and usually corresponds to more severe mechanical clearance or friction anomalies.

[0042] The correlation dimension was estimated using the Grassberger-Procaccia method: ; in, To reconstruct the first phase space A vector, To reconstruct the first phase space A vector, The number of trajectory points included in the statistics. For unit step function, Distance threshold For the distance between points on the trajectory is less than proportion, To be within the scale-invariant interval right The fitted slope.

[0043] To be within the scale-invariant interval right The slope of the fit reflects the geometric complexity of the attractor.

[0044] The volume shrinkage rate is estimated by the change in the trajectory neighborhood volume with the evolution step size: ; in, This refers to the volume shrinkage rate. For Centered on The neighborhood formed by the nearest neighbor trajectory points, The number of trajectory points included in the statistics. For evolutionary step size, Let be the volume of the neighborhood.

[0045] For Centered on, by its The neighborhood formed by the nearest neighbor trajectory points, with radius The ball, this point is commonly used in engineering implementation. The convex hull formed by the nearest neighbors is used as a substitute; It reflects the volume expansion or contraction trend of the system attractor during the evolution process.

[0046] The above three features are concatenated to obtain the degenerate manifold feature vector: ; in, For the eigenvectors of the degenerate manifold, The maximum Lyapunov exponent. To be within the scale-invariant interval right The fitted slope, This represents the volume shrinkage rate.

[0047] Step S3: Based on the causal direction between parameters in the multi-parameter time series dataset, and using the working condition fingerprint vector as a condition variable to exclude false causality, construct a degenerate causal directed graph with each parameter as a node, causal relationship as a directed edge, and causal strength as the edge weight.

[0048] Step S3 includes the following steps: For each parameter time series in the multi-parameter time series dataset, detrending and variance normalization are performed using the action cycle as the segment unit to eliminate the influence of magnitude differences on causality tests. For any two normalized parameter sequences... , The conditional Granger causality test is used to determine the causal direction. Restricted and complete models are constructed: ; ; in, For parameter sequences, For parameter sequences, This represents the total number of coefficients for the independent variables in the restricted model. , The regression coefficients of the constrained model, , , The regression coefficients are those of the complete model. The lag order of the regression model. For the condition variable vector, For the residuals of the constrained model, The residuals are for the complete model.

[0049] For the condition variable vector, the medium pipeline pressure is taken from the operating condition fingerprint vector. Opening command value and ambient temperature It is used to eliminate false causal relationships caused by the simultaneous fluctuation of two parameters due to changes in operating conditions.

[0050] Let the sums of squared residuals of the two models be respectively , Construct statistics: ; in, For the reason point to Statistical measure For the squared residuals of the constrained model, The squared residuals of the complete model, The dimension of the condition variable. This represents the number of samples used in the test.

[0051] when Exceeding the significance level When the corresponding critical value is reached, it is determined that there exists a... point to To determine the causal relationship, corresponding directed edges are constructed in the degenerate causal directed graph, with the direction of the edge representing the direction of the causal relationship. The causal relationships corresponding to each significant causal relationship are then mapped. The statistics are normalized over all edges in the graph and used as the edge weights of the corresponding directed edges. .

[0052] An adaptive adjustment mechanism is set up to determine the significance threshold. Under the initial healthy operating conditions of the electric actuator, the health baseline values ​​of the three components of the degenerate manifold eigenvector are calibrated. , , And at the same time, statistics on the benchmark within the periodic period. Standard deviation Define the percentage deviation of the maximum Lyapunov index from the healthy baseline value in the current action cycle: ; in, This represents the percentage deviation of the maximum Lyapunov index from a healthy benchmark. The maximum Lyapunov exponent. This is the health benchmark for the maximum Lyapunov index.

[0053] The adaptively adjusted significance threshold is: ; in, The significance level is the one commonly used in statistics. This is the sensitivity coefficient. for standard deviation This is the health benchmark for the maximum Lyapunov index.

[0054] The sensitivity coefficient is determined by the ratio of the mean to the standard deviation of the maximum Lyapunov index within the health baseline calibration period. This ratio reflects the stability of the index under healthy conditions: the more stable the baseline data... The smaller, The larger the threshold, the more pronounced the decrease, and the stronger the ability to capture weak causal relationships; the greater the fluctuation of the baseline data, the more pronounced the decrease. The smaller the value, the less likely it is to cause misjudgments due to noise in the reference itself. The larger the value, the more severely the current maximum Lyapunov index deviates from the healthy benchmark; the corresponding threshold is... The lower the value, the greater the sensitivity to capturing potential causal relationships when the system shows obvious signs of degradation.

[0055] Step S4: Perform topological analysis on the degenerate causal directed graph and calculate the topological degradation index vector representing the characteristics of degradation propagation.

[0056] Step S4 includes the following steps: In degenerate causal directed graph middle, It is a set of nodes identified by various parameters. Given a set of directed edges, identify the node with an in-degree of zero and the largest out-degree as the source node of degradation. This node represents the parameter channel where degradation first occurs. Starting from a node, use the directed graph shortest path algorithm to calculate the path to all end nodes with an out-degree of zero. The shortest directed path length (in terms of the number of edges), taken as the average, is used as the propagation path length: ; in, For the propagation path length, for To the end node The number of edges in the shortest directed path. This represents the total number of end nodes. A larger value indicates that the degradation propagates from the source to the end through more intermediate stages. Calculate the set of downstream reachable nodes from the degradation source node. Define branch diffusion: ; in, For branching diffusion, The total number of nodes in the degenerate causal directed graph. This represents the number of reachable nodes. A larger value indicates a wider range of degradation impact. A graph search method is used to detect the number of directed loops in the graph. Based on the total number of sides in the graph After normalization, the loop risk index is obtained: ; in, This is the loop risk index. The number of directed loops, This represents the total number of sides.

[0057] A value greater than zero indicates the presence of a degenerate positive feedback loop in the graph. The degenerate source node identifier, propagation path length, branch diffusion degree, and loop risk index are combined into a topology degradation index vector: ; in, This is a vector of topological degradation indices. For the propagation path length, For branching diffusion, This is the loop risk index.

[0058] Step S5: Using the physical mechanism of the degradation of the electric actuator as a priori constraint, and the characteristic vector of the degraded manifold and the topological degradation index vector as observational evidence, determine the dynamic weight vector of each parameter's contribution to health through Bayesian inference.

[0059] Step S5 includes the following steps: Based on the degradation physical mechanism of electric actuators, a priori degradation propagation order relationships are established for each parameter channel: abnormal rise in motor winding temperature precedes output shaft torque fluctuation, output shaft torque fluctuation precedes valve stem displacement deviation, and valve stem displacement deviation precedes abnormal drive current. The parameter channels involved in the fusion are denoted as... These correspond to the motor winding temperature, torque, displacement, and current path, respectively, and their dynamic weight vectors are denoted as... Each component is non-negative and sums to one.

[0060] To encode prior order relations as prior distribution parameters, the prior order conformity degree of each channel is defined using the edge weights between channels in the degenerate causal directed graph: ; in, For channel Prior order conformity For degenerate causal directed graphs with channels Pointing to channel The right to the side.

[0061] If the a priori order is satisfied, the direction of the directed edge is consistent with the propagation order specified by the aforementioned physical mechanism; if the a priori order is violated, the direction is opposite. The larger the value, the more the observed causal graph structure supports the channel. It is at a relatively advanced stage in the degeneration and propagation process.

[0062] The prior distribution with the Dirichlet distribution as the weight vector has the following concentration parameters: ; in, For the prior concentration parameter, This is the scaling factor. For channel Prior order conformity For all The nodes are taken as their arithmetic mean.

[0063] Take all The concentration parameter is the reciprocal of the mean of channels with positive values, allowing the scale of the concentration parameter to adapt to the edge weight distribution of the current actual causal graph. Channels with larger prior concentration parameters receive higher initial weights when observational evidence is insufficient.

[0064] Using the eigenvectors of degenerate manifolds and topological degradation index vector As current observational evidence, the strength of evidence for each channel's deviation from the health baseline is calculated. For each channel... Take the standardized deviation of the corresponding feature relative to the health benchmark. (For example, the temperature channel of the motor winding is taken as the standardized deviation of the temperature relative to the reference, and the torque, displacement, and current channels are taken as the standardized deviation of the manifold or topological features closely related to them relative to the reference.) After normalization, the strength of evidence is obtained: ; in, For the first The strength of evidence regarding the passage For channel The standardized deviation of the corresponding feature relative to the health baseline. For the first The sum of the absolute values ​​of the standardized deviations of each channel.

[0065] The posterior concentration parameters were obtained using the Dirichlet-multiple conjugate update rule: ; in, For posterior concentration parameters, For the prior concentration parameter, This represents the equivalent sample size for the evidence from this observation. For the eigenvectors of the degenerate manifold, This is a vector of topological degradation indices. To calculate the dimension of a vector, For the first The strength of evidence regarding the passage.

[0066] The components of the dynamic weight vector are obtained from the expectation formula of the posterior Dirichlet distribution: ; in, For each component of the weight vector, For posterior concentration parameters, To The posterior concentration parameters of each channel are summed through iteration.

[0067] When the loop risk index in the topology degradation index vector When the value is greater than zero, a penalty decay is applied to the weights of each channel involved in the directed loop: ; in, As a penalty decay factor, To form the set of edges for a directed cycle, For degenerate causal directed graphs with channels Pointing to channel The right to the border, It is the average of the edge weights of the loop.

[0068] The average weight of each edge in the loop reflects the strength of the causal relationship within the loop: the stronger the relationship and the higher the proportion of loops, the more significant the attenuation, thus preventing positive feedback loops from excessively amplifying the weights of related channels. After attenuation, all channel weights are renormalized to obtain the final dynamic weight vector. .

[0069] Step S6: Based on the dynamic weight vector, the degenerate manifold feature vector and the topology degradation index vector are weighted and fused, and combined with the cumulative number of actions in the working condition fingerprint vector, the comprehensive health of the electric actuator is calculated. The degradation level is determined according to the comprehensive health and the corresponding maintenance strategy is generated.

[0070] Step S6 includes the following steps: Based on the relative deviations of each component of the degenerate manifold eigenvector from the health baseline, the health sub-score of the degenerate manifold is calculated: ; in, For healthy sub-scores of degenerate manifolds, The maximum Lyapunov exponent. The maximum Lyapunov index is the health benchmark value. For the correlation dimension, As a health benchmark for correlation dimensions, This refers to the volume shrinkage rate. This serves as a healthy benchmark for volume shrinkage rate.

[0071] The score represents the health sub-score of the degenerate manifold; the smaller the value, the more severe the degradation reflected by that dimension.

[0072] The corresponding topological benchmark value is obtained based on the propagation path length, branch diffusion degree, loop risk index, and health benchmark calibration period in the topological degradation index vector. , The degree of deviation is used to calculate the causal topological health sub-score: ; in, For the causal topological health sub-score, For the length of the causal propagation path, The length of the transmission path under a healthy baseline. For causal branch diffusion, As a healthy benchmark for branching diffusion, This is the loop risk index.

[0073] The parameter channels participating in the fusion are denoted as These correspond to the motor winding temperature, torque, displacement, and current path, respectively, and their dynamic weight vectors are denoted as... Each component is non-negative and sums to one.

[0074] The dynamic weight vectors are grouped into two groups according to their physical meaning: the torque, displacement, and current channels correspond to the original components upon which the degenerate manifold features depend, and the sum of their weights serves as the manifold group weight; the motor winding temperature channel, as the channel containing the most frequently identified degenerate source nodes in the degenerate causal directed graph, has its weight used as the topology group weight. ; in, For the dynamic weights of the manifold group, For the dynamic weights of the topology group, These are the weighted components of the torque channel. These are the weighted components of the displacement channel. These are the weighted components of the current path. These are the weighted components of the temperature channel.

[0075] because The sum of all components is one, therefore The overall health score is obtained by weighting and summing the two health sub-scores using two sets of weights: ; in, For overall health, For the dynamic weights of the manifold group, For healthy sub-scores of degenerate manifolds, For the causal topological health sub-score, For the dynamic weights of the topology group.

[0076] The cumulative number of actions in the working condition fingerprint vector Lifespan correction is applied to the overall health status. The rated lifespan of the electric actuator is denoted as [number of operations]. (Rated values ​​given on the equipment nameplate or in the selection design), calculate the life cycle ratio. In the historical maintenance and failure record database of the same model of equipment, the proportion of life cycles corresponding to the failure events that have occurred is statistically analyzed, and the mean value is calculated. with standard deviation This information is continuously updated as database records accumulate, and the initial correction ratio is determined accordingly. ; in, To correct the initial ratio, This represents the mean of the sample of lifecycle proportions corresponding to the failure event. This represents the standard deviation of the proportion of lifecycles corresponding to the failure event.

[0077] when No correction is applied at times; when When applying a decreasing correction factor: ; in, This represents the lifespan ratio. The decay rate coefficient, To correct the initial ratio, To decrease the correction factor, For overall health status.

[0078] The decay rate coefficient is given by... The correction coefficient is determined by self-reflection. It decays sufficiently when it approaches one (the life cycle ratio approaches saturation).

[0079] The threshold for determining the level of deterioration is also adaptively determined based on statistical data from the health benchmark period. The mean of the overall health status sample within the benchmark period is denoted as... Standard deviation is The boundaries for dividing the five levels are as follows: ; in, This represents the mean of the overall health score sample. The standard deviation of the overall health sample. , , , This represents the health threshold boundary.

[0080] when The time was determined to be healthy; It was initially determined to be a minor degeneration; It was determined to be moderate degradation at that time; It was determined to be severely degraded at that time; A failure risk is identified at a certain point. Boundary values ​​are all derived from the specific equipment's own health baseline data, which can adapt to the individual differences of electric actuators under different models and operating conditions. Corresponding maintenance strategies are generated for different degradation levels. For example, for the health level, regular inspection cycles are maintained; for the minor degradation level, the inspection frequency is increased and trends are recorded; for the moderate degradation level, preventive maintenance is arranged; for the severe degradation level, shutdown maintenance is arranged; and for the failure risk level, emergency shutdown and replacement recommendations are triggered.

[0081] Based on overall health Numerical sequence over multiple consecutive action cycles As input, the grey prediction model GM(1,1) is used for extrapolation prediction. An accumulated generation sequence is constructed. Establish the grey differential equation ,in, , The parameters to be estimated are solved using the least squares method: ; in, For development coefficient, The amount of gray action, To design the matrix, This is the first-order cumulative reduction of the original sequence.

[0082] , Constructed from the accumulated sequence using the standard GM(1,1) modeling method. The elements are the original sequence. ( ), Each row consists of the mean of adjacent accumulated values ​​and a constant term.

[0083] Prediction formula: ; in, For the accumulating sequence One predicted value, To accumulate the first value of the generated sequence, These are the predicted values ​​for the original sequence. To generate a cumulative sequence, For the accumulating sequence One predicted value, For periodic indexes, For development coefficient, This represents the amount of gray action.

[0084] Based on the above prediction sequence, the threshold for the first time the overall health level drops to the failure risk level is calculated. The following required number of remaining operating cycles is used as an estimate of the remaining usable life. .Will With preset safety margin period Compare, The number of lead times required for maintenance of the same type of equipment in history is determined and continuously updated; when When this happens, the current degradation level will be increased by one level, and a corresponding upgrade of the maintenance strategy will be triggered. The cumulative number of actions in the working condition fingerprint vector By linking these parameters, the full lifecycle degradation curve of the electric actuator can be plotted and output, providing a basis for equipment operation and maintenance scheduling.

[0085] Example 2: Figure 2 This is an architecture diagram of the multi-parameter fusion electric actuator health assessment system of the present invention. The system includes: The multi-parameter acquisition module is used to acquire multi-source parameters of the electric actuator, obtain valve load condition context information and encode it into a condition fingerprint vector, and perform time-series alignment of the multi-source parameters to obtain a multi-parameter time-series dataset. The phase space reconstruction module is used to extract three parameters—torque, displacement, and current—from the multi-parameter time series dataset. Combined with the valve type encoding in the working condition fingerprint vector, it constructs a three-dimensional coupled phase space and performs phase space reconstruction to extract the degenerate manifold feature vector of the electric actuator. The causal graph construction module is used to construct a degenerate causal directed graph based on the causal direction between parameters in a multi-parameter time series dataset, and to exclude false causality by using the working condition fingerprint vector as a condition variable. The graph has each parameter as a node, causal relationship as a directed edge, and causal strength as the edge weight. The degradation feature analysis module is used to perform topological structure analysis on the degradation causal directed graph and calculate the topological degradation index vector representing the degradation propagation characteristics. The contribution weight analysis module is used to determine the dynamic weight vector of each parameter's contribution to health by using the degradation physical mechanism of the electric actuator as a prior constraint, the degradation manifold feature vector and the topological degradation index vector as observational evidence, and Bayesian inference. The maintenance decision generation module is used to perform weighted fusion of degraded manifold feature vector and topological degradation index vector based on dynamic weight vector, and combine it with the cumulative number of actions in the working condition fingerprint vector to calculate the comprehensive health of electric actuators. Based on the comprehensive health, the degradation level is determined and the corresponding maintenance strategy is generated.

[0086] Example 3: The performance of the present invention was compared with that of existing methods.

[0087] This embodiment uses three electric actuators of different valve types in a petrochemical pipeline system as reference nodes, namely Device 1 (gate valve, valve type code). Device 2 (ball valve) Device 3 (butterfly valve) This study verifies the applicability of the present invention in scenarios with multiple valve types and makes a quantitative comparison with two existing technologies that are similar to the technical route of the present invention: the first existing technology is a fixed-weight multi-parameter health assessment method based on principal component analysis (PCA), and the second existing technology is a health trend prediction method based on long short-term memory network (LSTM).

[0088] Table 1. Operating Condition Fingerprint Vectors and Basic Parameters for Three Valve Types

[0089] The following uses device two (ball valve) as an example for the experiment, and the key intermediate results of the other two devices are given for comparison.

[0090] Number of sampling points in the second operating cycle of the device Sampling time interval The output shaft torque is obtained. Valve stem displacement Drive current As shown in the table below: Table 2 Sampling data for a single action cycle of Device 2 (ball valve)

[0091] According to the formula Calculation: The differences between adjacent torques are 2, 3, 4, 5, 6, 7, 8, 9, 10. After calculation .

[0092] Table 3 Comparison of Torque Variation Rate and Adaptive Search Interval for Three Valve Types

[0093] This invention adaptively scales the search range of the optimal delay time to suit the torque variation characteristics of different valve types, thus avoiding feature distortion caused by uniform parameters.

[0094] Within the adjusted search interval, device 2 calculates the optimal delay time using the mutual information method and the Cao method. , , (Based on the number of sampling points), optimal embedding dimension The trajectory sequence was reconstructed using this method, and the three degenerate manifold features were calculated according to the formula. The degenerate manifold feature vector of device 2 was then compared with the healthy baseline as follows: Table 4. Phase space reconstruction results and degenerate manifold eigenvectors of Device 2 (ball valve)

[0095] Right now .

[0096] For temperature channels Torque Channel Displacement Channel Current path Perform conditional Granger causality tests pairwise (condition variable selection) , , ), at the critical value The following test identified four significant causal edges: Table 5. Results of edge weights and significance tests for the degenerate causal directed graph of Device 2 (ball valve)

[0097] according to ,Pick , Adaptive threshold The value was significantly lower than the conventional significance level of 0.05, reflecting the weak causal capture ability enhanced by the significant deviation of the maximum Lyapunov index from the baseline.

[0098] Nodes in the diagram , A node with zero in-degree is considered a degenerate source node. The end node is (Out-degree is zero), shortest path Since there are 3 edges, Downstream reachable node set , The diagram contains loops. , Total number of sides , .

[0099] Table 6. Topology Degradation Index Vector for Device 2 (Ball Valve)

[0100] Prior order conformity: ( (Conforms to prior order) ( (Conforms to prior order) ( conform to, violation), Take the average of positive values. Scaling factor .

[0101] Table 7 Dynamic Weighted Bayesian Fusion Process of Device 2 (Ball Valve)

[0102]

[0103] in, Loop penalty coefficient , ,right , After applying this attenuation to both channels and renormalizing, the final dynamic weight vector is obtained: ; Degenerate manifold health sub-score: ; Causal topological health sub-score: ; manifold weights Topology group weights Overall health status: ; Lifetime correction: ,Pick ,because No corrections are applied. .

[0104] Table 8 Comparison of Comprehensive Health and Degradation Level Judgment Results for Three Valve Types

[0105] Among them, the ratio of device life cycle to three factors Exceeding the initial correction ratio ,according to calculate: ; Before the correction, device 3 was judged to be severely degraded; after the correction, it was judged to be at risk of failure. This reflects the necessity of adjusting the overall health status based on the cumulative number of operations when the device is nearing its rated life.

[0106] Due to the degradation of the source node of device two The (motor winding temperature channel) clearly points to the primary anomaly of the electromagnetic and heat dissipation subsystems. Based on this, the maintenance strategy generated by the system prioritizes checking the insulation status of the motor windings and the heat dissipation channel, and conducts a coordinated investigation of the downstream torque and displacement transmission chains. Compared with existing technologies that only output health values, this has a stronger root cause orientation.

[0107] Take the historical comprehensive health data of Device 2 from four consecutive monitoring windows (each window covering 2000 action cycles). (correspond Establish a GM(1,1) model for iterations of 24000, 26000, 28000, and 30000. The cumulative sequence is... , , , The development coefficient is obtained by solving the least squares method. Ash action amount The prediction formula is: ; Table 9 Comparison of predicted and measured values ​​of colorimetric parameters for Device 2 (ball valve)

[0108] It can be seen that, in At this point, the comprehensive health score (0.3594) obtained by the present invention based on causal topology and manifold characteristics is significantly lower than the predicted value (0.4765) extrapolated from the disaster color trend, indicating that the equipment degradation rate has accelerated compared to the historical trend, and this value has fallen below the failure risk threshold. Based on this, the remaining usable lifespan is determined. The time, less than the preset safety margin The second instance triggered a maintenance strategy upgrade to immediate shutdown and repair. This result verifies that the present invention can capture accelerated degradation phenomena more promptly compared to pure trend extrapolation methods (such as prior art 2).

[0109] A quantitative comparison with existing technologies.

[0110] The monitoring data of the same batch of three devices were processed using existing technology one (PCA fixed weight method) and existing technology two (LSTM trend prediction method), respectively. The actual degradation degree obtained from the subsequent disassembly and inspection of the back rack (corresponding to the actual health level: Device 1 0.71, Device 2 0.34, Device 3 0.26) was used as a benchmark to calculate the evaluation error of each method. Simultaneously, the time when anomalies were discovered during manual offline inspection (Diagram showing the time when anomalies were discovered for Device 2) was used as the benchmark. (times) and the actual number of times the equipment was scrapped (corresponding to device two) Using the actual remaining lifetime (RUL) of 2000 cycles as a benchmark, the early detection lead time and RUL prediction error are calculated.

[0111] Table 10 Comparison of the evaluation effects of the present invention and two existing technologies

[0112] As shown in Table 10, the root mean square error (RMSE) of the health assessment of the present invention on three different valve types (0.0153) is significantly lower than that of the LSTM trend prediction method (0.0408) and the PCA fixed weight method (0.0851). On device two, the present invention issues an early warning 2500 action cycles earlier than manual offline inspection, significantly better than the 1300 cycles of the LSTM method and the 700 cycles of the PCA method. In predicting remaining usable life, the present invention has a prediction error of 0, while the LSTM method has an error of 1200 times, and the PCA fixed weight method, lacking causal propagation and dynamic weighting mechanisms, can only provide a rough extrapolation with an error of 3000 times. These quantitative results correspond to the beneficial effects described in the invention's description, such as early sensitivity, root cause localization capability, dynamic weight stability, and trend prediction accuracy, verifying the effectiveness and advancement of the present invention's technical solution.

[0113] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for assessing the health of an electric actuator using multi-parameter fusion, characterized in that, The method includes: Step S1: Collect multi-source parameters of the electric actuator, and at the same time obtain the valve load condition context information and encode it into a condition fingerprint vector. Perform time-series alignment on the multi-source parameters to obtain a multi-parameter time-series dataset. Step S2: Extract three parameters, torque, displacement and current, from the multi-parameter time series dataset. Combine the valve type encoding in the working condition fingerprint vector to construct a three-element coupled phase space and reconstruct the phase space. Extract the degenerate manifold feature vector of the electric actuator. Step S3: Based on the causal direction between parameters in the multi-parameter time series dataset, and using the working condition fingerprint vector as a condition variable to exclude false causality, construct a degenerate causal directed graph with each parameter as a node, causal relationship as a directed edge, and causal strength as the edge weight. Step S4: Perform topological analysis on the degenerate causal directed graph and calculate the topological degradation index vector representing the characteristics of degradation propagation; Step S5: Using the degradation physical mechanism of the electric actuator as a priori constraint, and the degradation manifold feature vector and topological degradation index vector as observational evidence, determine the dynamic weight vector of each parameter's contribution to health through Bayesian inference. Step S6: Based on the dynamic weight vector, the degenerate manifold feature vector and the topology degradation index vector are weighted and fused, and combined with the cumulative number of actions in the working condition fingerprint vector, the comprehensive health of the electric actuator is calculated. The degradation level is determined according to the comprehensive health and the corresponding maintenance strategy is generated.

2. The method for assessing the health of electric actuators using multi-parameter fusion according to claim 1, characterized in that, Step S1 specifically includes: Multi-source parameters include electrical parameters, mechanical parameters, and state parameters. Electrical parameters include drive current, terminal voltage, and power factor; mechanical parameters include output shaft torque, valve stem displacement, and angular velocity; state parameters include motor winding temperature, reducer oil temperature, and housing vibration acceleration. Valve load condition context information includes medium pipeline pressure, ambient temperature, valve type code, opening command value, and cumulative number of actions; The entire stroke from receiving the opening command to reaching the target opening degree of the electric actuator is taken as an action cycle. Within each action cycle, the sampling points of each parameter are synchronized and aligned with a unified time reference. Each action cycle in the multi-parameter time series dataset is associated with the corresponding working condition fingerprint vector.

3. The method for assessing the health of electric actuators using multi-parameter fusion according to claim 1, characterized in that, The specific steps of constructing the ternary coupled phase space and reconstructing the phase space in step S2 include: Using the output shaft torque, valve stem displacement, and drive current in the multi-parameter time series dataset as three original components, a ternary coupled state vector is constructed. The optimal delay time between each component in the ternary coupled state vector is determined by the mutual information method, and the optimal embedding dimension of the phase space is determined by the Cao method. The phase space of the ternary coupled state vector is reconstructed using the optimal delay time and the optimal embedding dimension to obtain the sequence of trajectory points in the reconstructed phase space. Based on the valve type code in the working condition fingerprint vector, the torque-displacement coupling time delay characteristics corresponding to the current valve type are determined, and the search range of the optimal delay time is adaptively adjusted accordingly.

4. The method for assessing the health of electric actuators using multi-parameter fusion according to claim 1, characterized in that, The extraction of the degenerate manifold feature vector of the electric actuator in step S2 specifically includes: On the trajectory point sequence, the maximum Lyapunov exponent, the correlation dimension of the reconstructed phase space, and the volume shrinkage rate of the reconstructed phase space are calculated, and the three are concatenated to obtain the degenerate manifold feature vector.

5. The method for assessing the health of an electric actuator using multi-parameter fusion according to claim 1, characterized in that, In step S3, the construction of the degenerate causal directed graph specifically includes: For each parameter time series in the multi-parameter time series dataset, detrending and variance normalization are performed using the action cycle as the segment unit. For any two normalized parameter time series, the conditional Granger causality test is used to determine the causal direction relationship between them. The conditional Granger causality test uses the medium pipeline pressure, opening command value and ambient temperature in the operating condition fingerprint vector as condition variables to eliminate false causal associations caused by changes in operating conditions. When the causal relationship between two parameters passes the significance test, a corresponding directed edge is established in the degenerate causal directed graph, and the direction of the directed edge is the causal action direction; the F-statistic obtained from the conditional Granger causality test is normalized and used as the edge weight of the directed edge. The health baseline values ​​are reference values ​​for the maximum Lyapunov index, correlation dimension, and volume shrinkage rate calibrated under healthy conditions during the initial operation of the electric actuator. When the maximum Lyapunov exponent in the eigenvector of a degenerate manifold exceeds its corresponding health baseline, the decision threshold for significance testing is lowered to enhance the ability to capture weak causal relationships.

6. The method for assessing the health of an electric actuator using multi-parameter fusion according to claim 1, characterized in that, In step S4, the calculation process of the topology degradation index vector specifically includes: In a degenerate causal directed graph, the node with zero in-degree and the largest out-degree is identified as the degenerate source node. The degenerate source node represents the parameter channel where the degeneracy first occurs. Starting from the degenerate source node, calculate the shortest directed path length from it to all end nodes with an out-degree of zero, and take the average of all shortest directed path lengths as the propagation path length. The proportion of downstream reachable nodes of the degenerate source node to the total number of nodes in the degenerate causal directed graph is calculated as the branch diffusion degree. The number of directed loops in a degenerate causal directed graph is detected, and the normalized number of directed loops is used as the loop risk index. The identifier of the degradation source node, the propagation path length, the branch diffusion degree, and the loop risk index are combined into a topology degradation index vector.

7. The method for assessing the health of an electric actuator using multi-parameter fusion according to claim 1, characterized in that, In step S5, the physically constrained Bayesian dynamic weight fusion specifically includes: A priori order relationship is established based on the degradation physical mechanism of electric actuators: abnormal rise in motor winding temperature precedes output shaft torque fluctuation, output shaft torque fluctuation precedes valve stem displacement deviation, and valve stem displacement deviation precedes abnormal drive current. The priori order relationship is encoded as prior distribution parameters of a dynamic weight vector. The expected value of the prior distribution parameters decreases monotonically from front to back along the priori order relationship, so that the initial weight corresponding to the parameter ranked earlier in the priori order relationship is greater than the initial weight corresponding to the parameter ranked later. Using the degenerate manifold feature vector and topological degradation index vector as current observation evidence, Bayesian update is performed on the prior distribution parameters to obtain the posterior distribution. The expected value is taken from the posterior distribution as the final weight of each parameter in the dynamic weight vector. When the loop risk index in the topology degradation index vector is greater than zero, a penalty decay factor is applied to the weights corresponding to each parameter involved in the directed loop to suppress the weight concentration caused by the positive feedback loop.

8. The method for assessing the health of an electric actuator using multi-parameter fusion according to claim 1, characterized in that, In step S6, the calculation of overall health and the determination of the degree of degradation specifically include: The health sub-score of the degenerate manifold is calculated based on the degree of deviation between each topological invariant in the eigenvector of the degenerate manifold and the health baseline value. Based on the degree of deviation between the propagation path length, branch diffusion degree, and loop risk index in the topology degradation index vector and the topology benchmark value corresponding to the health benchmark value, the causal topology health sub-score is calculated. The overall health score is obtained by weighting and summing the health scores of the degenerate manifold and the causal topological health sub-scores using a dynamic weight vector. The overall health score is corrected based on the cumulative number of actions in the working condition fingerprint vector: when the cumulative number of actions exceeds the preset life cycle ratio, a decreasing correction coefficient is applied to the overall health score. Based on the revised comprehensive health value range, electric actuators are classified into five degradation levels: healthy, slightly degraded, moderately degraded, severely degraded, and at risk of failure, and corresponding maintenance strategies are generated.

9. The method for assessing the health of an electric actuator using multi-parameter fusion according to any one of claims 1 to 8, characterized in that, The prediction of degradation trends specifically includes: Using the numerical sequence of comprehensive health score over multiple consecutive action cycles as input, a grey prediction model is used to extrapolate and predict the future evolution trend of comprehensive health score. Based on the prediction results, the number of remaining operating cycles required for the overall health status to drop below the threshold corresponding to the failure risk level for the first time is calculated as the estimated remaining usable life of the electric actuator. When the estimated remaining usable life is less than the preset safety margin period, the degradation level will be increased by one level and the maintenance strategy will be upgraded. The remaining usable life estimate is correlated with the cumulative number of actions in the operating condition fingerprint vector to generate the full life cycle degradation curve of the electric actuator.

10. A health assessment system for electric actuators based on multi-parameter fusion, characterized in that, The system includes: The multi-parameter acquisition module is used to acquire multi-source parameters of the electric actuator, obtain valve load condition context information and encode it into a condition fingerprint vector, and perform time-series alignment of the multi-source parameters to obtain a multi-parameter time-series dataset. The phase space reconstruction module is used to extract three parameters—torque, displacement, and current—from the multi-parameter time series dataset. Combined with the valve type encoding in the working condition fingerprint vector, it constructs a three-dimensional coupled phase space and performs phase space reconstruction to extract the degenerate manifold feature vector of the electric actuator. The causal graph construction module is used to construct a degenerate causal directed graph based on the causal direction between parameters in a multi-parameter time series dataset, and to exclude false causality by using the working condition fingerprint vector as a condition variable. The graph has each parameter as a node, causal relationship as a directed edge, and causal strength as the edge weight. The degradation feature analysis module is used to perform topological structure analysis on the degradation causal directed graph and calculate the topological degradation index vector representing the degradation propagation characteristics. The contribution weight analysis module is used to determine the dynamic weight vector of each parameter's contribution to health by using the degradation physical mechanism of the electric actuator as a prior constraint, the degradation manifold feature vector and the topological degradation index vector as observational evidence, and Bayesian inference. The maintenance decision generation module is used to perform weighted fusion of degraded manifold feature vector and topological degradation index vector based on dynamic weight vector, and combine it with the cumulative number of actions in the working condition fingerprint vector to calculate the comprehensive health of electric actuators. Based on the comprehensive health, the degradation level is determined and the corresponding maintenance strategy is generated.

Images