Real-time monitoring and early warning system for electromechanical system faults based on digital twin
By combining digital twin technology and an improved GDN model, and utilizing dual perturbation response and curvature delay mapping, the problem of early degradation identification and impact range judgment in complex electromechanical systems was solved, enabling accurate and timely monitoring and early warning of electromechanical system faults.
Patent Information
- Application Number
- CN202611465596.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-20
- Publication Date
- 2026-11-17
AI Technical Summary
Existing digital twin fault monitoring systems struggle to distinguish between normal operating point migration and component local dynamic degradation, lacking the ability to identify early degradation and determine its impact range. In particular, they lack structural change analysis of neighborhood response relationships under opposite-direction disturbances in complex electromechanical systems.
By combining digital twin technology and an improved GDN model, early degradation of electromechanical components can be identified through dual perturbation response, curvature delay mapping, and mirror neighbor sequence splitting analysis. Dynamic response characteristics and mirror neighbor sequence splitting data can be generated to achieve accurate fault identification and timely early warning.
It improves the accuracy and adaptability of early fault monitoring in electromechanical systems, effectively characterizes gap increase, loosening, friction changes and asymmetric stiffness changes, and enhances the timeliness of fault location and real-time early warning.
Smart Images

Figure CN123113587A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of digital twin and electromechanical system fault monitoring technology, and in particular to a real-time monitoring and early warning system for electromechanical system faults based on digital twin. Background Technology
[0002] With the increasing automation of industrial equipment, complex electromechanical coupling relationships are formed between motors, bearings, couplings, gear transmission mechanisms, and control units. The operating status of equipment is affected by load, control commands, component wear, and the operating environment. Existing electromechanical system fault monitoring mainly relies on vibration, current, temperature, speed, and control status for condition identification. Digital twin technology can establish a state mapping between physical equipment and virtual models, and combine real-time operating data to perform state synchronization, fault analysis, and operational early warning, providing a technical foundation for online monitoring of complex electromechanical systems.
[0003] Current digital twin fault monitoring primarily identifies anomalies based on deviations between real-time monitored values, twin predicted values, and historical states. However, monitoring results are easily affected by changes in operating conditions, making it difficult to distinguish between normal operating point migration and localized component dynamic degradation. Graphical models built based on sensor correlations mainly reflect statistical associations between historical data, lacking dynamic characterization of response gain, response curvature, response delay, and recovery process after controlled disturbances to electromechanical components. This makes it difficult to detect early degradation caused by increased clearance, loosening, friction changes, and asymmetric stiffness variations in a timely manner. Existing early warning methods lack analysis of structural changes in neighborhood response relationships under disturbances in opposite directions, resulting in insufficient ability to identify early degradation and determine its impact range before fault characteristics have significantly exceeded limits.
[0004] Therefore, how to provide a real-time monitoring and early warning system for electromechanical system faults based on digital twins is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] One objective of this invention is to propose a real-time monitoring and early warning system for electromechanical system faults based on digital twins. This invention combines digital twin technology and an improved GDN model, and identifies early degradation of electromechanical components through dual perturbation response, curvature delay mapping, and mirror neighbor sequence splitting analysis. It has the advantages of accurate fault identification, strong adaptability to operating conditions, and timely early warning.
[0006] According to an embodiment of the present invention, a real-time monitoring and early warning system for electromechanical system faults based on digital twins includes: The data processing module is used to collect operating data of the electromechanical system and perform alignment processing to generate standard electromechanical status data; The twin synchronization module is used to construct a digital twin based on standard electromechanical status data and component connection relationships, synchronize the current state, and generate twin synchronization status data. The perturbation response module is used to apply local virtual perturbations in opposite directions to the component under analysis based on twin synchronization state data, collect the responses of related components, and generate dual perturbation response fringes. The curvature delay mapping module is used to extract response gain, response curvature, response delay and recovery features based on dual perturbation response ripples, and generate bidirectional curvature delay maps; The response analysis module is used to input the bidirectional curvature delay map into the improved GDN model, which has a dual perturbation response coding layer, a curvature delay map learning layer and a mirror neighbor split readout layer, to generate dynamic response features and mirror neighbor split data. The anomaly screening module is used to identify changes in local response relationships based on dynamic response characteristics and generate candidate anomaly components. The degradation confirmation module is used to generate component degradation status data based on the neighbor set and response sorting under opposite perturbation of candidate abnormal components and mirror neighbor order split data. The fault early warning module is used to determine the duration and scope of degradation based on component degradation status data and continuous cycle dynamic response characteristics, and to generate real-time fault monitoring and early warning results.
[0007] Optionally, the electromechanical system operation data includes electromechanical operation status data, operation control condition data, electromechanical component configuration data, and historical operation monitoring data.
[0008] Optionally, generating standard electromechanical condition data includes: Read electromechanical operating status data, operating control condition data, electromechanical component configuration data, and historical operating monitoring data, and establish data association relationships according to equipment identification, component identification, and data acquisition time; Based on data association, the operating status, control status and operating conditions of each electromechanical component are extracted and arranged according to the time of collection to form a component status sequence. The component status sequence is processed by filtering out abnormal records, filling in missing records, aligning sampling periods, and unifying units to generate preprocessed electromechanical status data. By associating preprocessed electromechanical status data with electromechanical component configuration data according to equipment and component identifiers, a correspondence between component status, component connection, and operating conditions is established, and standard electromechanical status data is generated.
[0009] Optionally, generating twin synchronization state data includes: Read component status, control status, operating conditions, component configuration parameters, and component connection relationships from standard electromechanical condition data, and establish a mapping relationship between physical components and twin components according to component identifiers; Establish twin component nodes based on component configuration parameters, connect twin component nodes according to component connection relationships, write control status and operating conditions into the corresponding twin component nodes, and construct a digital twin; Extract normal operation records that match the current operating conditions from historical operation monitoring data, calculate the state deviation between the physical component state and the twin component state, adjust the corresponding twin parameters according to the parameter value range and parameter state response relationship until the state deviation enters the state fluctuation range formed by the normal operation records, and generate twin synchronized state data.
[0010] Optionally, generating the dual perturbation response ripple includes: Read the twin synchronization status data, extract adjustable twin parameters, parameter value ranges and component connection relationships from the electromechanical component configuration data, and determine the component to be analyzed according to the component identifier; The twin synchronization state of the component to be analyzed is copied, and the first direction perturbation state, the second direction perturbation state and the unperturbed state are established. The operating conditions, control states and parameters of the non-component to be analyzed are kept consistent. Local virtual perturbations with the same amplitude and opposite directions are applied to the same adjustable twin parameter. Run the first direction perturbation state, the second direction perturbation state, and the unperturbed state respectively. Record the reference response segment before perturbation is applied, the perturbation hold response segment, and the perturbation removal recovery response segment according to the same sampling period. Extract the associated components that generate state responses along the component connection relationship to form the first direction response trajectory, the second direction response trajectory, and the unperturbed response trajectory. Based on three types of response trajectories—identifier of the component to be analyzed, identifier of the associated component, identifier of the adjustable twin parameter, and alignment of the sampling time—a baseline response segment, a perturbation-preserved response segment, and a perturbation-removed recovery response segment are combined. A complete response acquisition process is defined as a monitoring cycle, generating a dual perturbation response pattern.
[0011] Optionally, generating the bidirectional curvature delay map includes: Read the dual perturbation response ripples, and extract the first direction response trajectory, the second direction response trajectory, and the unperturbed response trajectory according to the identifier of the component to be analyzed and the identifier of the associated component; For the same state variable, calculate the state change of the first direction response trajectory and the second direction response trajectory relative to the unperturbed response trajectory, and generate the first direction response gain and the second direction response gain according to the corresponding local virtual perturbation amplitude normalization. Generate the first direction response curvature and the second direction response curvature according to the response slope change of the two direction response trajectories at continuous sampling time. The baseline response segment is extracted from the unperturbed response trajectory. The response change boundary is determined by adding or subtracting three times the standard deviation of the state variable mean. The time interval from when the perturbation is applied to the first time the response of the associated component exceeds the response change boundary is determined as the response delay. The time interval from when the perturbation is removed to when the response of the associated component enters the response change boundary for three consecutive sampling times is determined as the recovery feature. Using electromechanical components as graph nodes and the response relationship between the component to be analyzed and related components as directed connections, the response gain, response curvature, response delay and recovery characteristics of each perturbation direction are written into the corresponding directed connections to construct the first direction curvature delay graph and the second direction curvature delay graph, and then aligned according to the component identifier to form the dual curvature delay graph.
[0012] Optionally, the generation of dynamic response features and mirror neighbor-order split data includes: Based on the GDN model, the original GDN graph structure learning method based on node embedding similarity to establish graph relationships is replaced by a dynamic response relationship learning method based on dual curvature time delay graph. A dual perturbation response encoding layer is added at the model input and a mirror neighbor order split readout layer is added at the model output to construct an improved GDN model. The dual curvature delay map is input into the dual perturbation response coding layer. According to the node characteristics and directed connection characteristics of the two directional maps of the component to be analyzed, the associated component and the state variable, directional pairing coding is performed on the response gain, response curvature, response delay and recovery characteristics to generate directional common response characterization and directional difference response characterization. The common directional response representation and the different directional response representation are input into the curvature delay graph learning layer. The directed connection in the dual curvature delay graph limits the range of neighborhood information transmission. Connection weights are generated based on response gain, response curvature, response delay and recovery features. Weighted aggregation of associated component representations and time-series prediction are performed to generate the current response representation and the predicted response representation and combine them to form dynamic response features. The dynamic response features are input into the mirror neighbor order split readout layer. The response gain sorting, response delay sorting, and response curvature sorting are performed on the associated components corresponding to the same component to be analyzed according to the first and second directions, respectively. The sets of associated components in the two directions are aligned, and the proportion of non-shared neighbors, the proportion of response gain sorting exchange, the proportion of response delay sorting exchange, and the proportion of response curvature sorting exchange are calculated to generate mirror neighbor order split data. Historical state records, operating conditions, control states, component configuration parameters, and component connection relationships corresponding to normal operating cycles are extracted from historical operation monitoring data, operation control condition data, and electromechanical component configuration data. Model training samples are generated through twin synchronization, local virtual perturbation, and dual curvature time delay mapping. The improved graph bias network (GDN) model is trained using dynamic response prediction error and mirror neighbor consistency error as joint loss. The dual curvature time delay map of the current monitoring cycle is input into the trained model, and the dynamic response characteristics and mirror neighbor split data are output.
[0013] Optionally, generating candidate anomaly components includes: Read the dynamic response characteristics and mirror adjacent sequence split data, and extract the response gain deviation, response curvature deviation, response delay deviation and recovery characteristic deviation according to the component identification; Normal operating cycles are extracted from historical operation monitoring data. Normal operating cycles that match the current operating conditions are selected based on operation control condition data. The 95th percentile values of various dynamic response deviations and mirror adjacent sequence split data are calculated to determine the local dynamic response judgment boundary. Components whose mirrored adjacent order split data exceeds the corresponding judgment boundary are formed into a first candidate set. Components whose response gain deviation, response curvature deviation, response delay deviation, and recovery feature deviation exceed the corresponding judgment boundary are formed into a second candidate set. The first and second candidate sets are merged and duplicate components are removed to generate candidate abnormal components.
[0014] Optionally, the generated component degradation state data includes: Read candidate abnormal components, continuously extract dynamic response features and mirror neighbor sequence split data according to the monitoring cycle, arrange response curvature deviation, response delay deviation, recovery feature deviation and mirror neighbor sequence split data according to component identifier, and generate candidate component degradation sequence; Calculate the characteristic change of adjacent monitoring cycles of candidate component degradation sequence, extract normal operation cycle from historical operation monitoring data, generate corresponding dynamic response characteristics and mirror adjacent sequence split data through response analysis module, calculate the 95th percentile value of normal cycle characteristic change, and determine the monitoring cycle that continuously exceeds the corresponding 95th percentile value as the degradation duration cycle. By associating candidate abnormal components, degradation duration, dynamic response deviation, and mirror neighbor sequence split data, component degradation status data is generated.
[0015] Optionally, the generation of real-time fault monitoring and early warning results includes: Read component degradation status data, extract component connection relationships from electromechanical component configuration data, and identify components in a continuous degradation state and directly connected components according to component identification; Dynamic response characteristics are continuously extracted according to the monitoring cycle to determine whether the dynamic response of directly connected components continues to exceed the limit, and the degradation impact range is determined layer by layer along the component connection relationship. A Level 1 warning is established when the degradation impact is limited to components in a state of continuous degradation; a Level 2 warning is established when the degradation impact extends to directly connected components; and a Level 3 warning is established when the degradation impact extends to connected components beyond the directly connected components. The system associates component identification, dynamic response deviation, mirror adjacent sequence split data, degradation duration, degradation impact range, and warning level to generate real-time fault monitoring and warning results.
[0016] The beneficial effects of this invention are: This invention proposes a real-time monitoring and early warning system for electromechanical system faults based on digital twins. It constructs a digital twin using standard electromechanical state data and applies opposite-direction local virtual perturbations to the component under analysis based on synchronized operating conditions. This forms a dual perturbation response pattern comprising a reference response segment, a perturbation-held response segment, and a perturbation-removed recovery response segment. Furthermore, it extracts response gain, response curvature, response delay, and recovery features to construct a dual curvature delay map. Compared to methods that directly rely on the absolute values of vibration, current, temperature, and rotational speed for fault diagnosis, this invention utilizes the local dynamic response relationship under controlled perturbations to describe the state changes of electromechanical components. This reduces the interference of operating condition changes on fault identification results, effectively characterizing early dynamic anomalies caused by increased clearance, loosening, friction changes, and stiffness degradation, thereby improving the accuracy and adaptability of early fault monitoring in electromechanical systems.
[0017] This invention transforms traditional graph relation learning based on node statistical association into dynamic relation learning based on local controlled responses by setting a dual perturbation response encoding layer, a curvature delay graph learning layer, and a mirrored neighbor order split readout layer in the GDN model. It also performs mirrored comparisons of the associated component sets, response gain ranking, response delay ranking, and response curvature ranking under opposite-direction perturbation conditions, forming mirrored neighbor order split data. By combining dynamic response deviation, mirrored neighbor order split changes, and continuous monitoring cycle degradation states, it can identify abnormal components with continuously changing local dynamic response relationships and determine the degradation impact range and early warning level along the component connection relationships. This expands fault early warning from identifying state quantity exceeding limits to identifying local dynamic structural changes, improving the detection capability of asymmetric degradation, local loosening, and early transmission characteristic anomalies, and enhancing the timeliness of fault location and real-time early warning. Attached Figure Description
[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1This is a system structure diagram of the real-time monitoring and early warning system for electromechanical system faults based on digital twins proposed in this invention; Figure 2 The flowchart shows the improved GDN model structure and data processing in the real-time monitoring and early warning system for electromechanical system faults based on digital twins proposed in this invention. Figure 3 This is a schematic diagram illustrating the construction process of the dual curvature time delay map in the real-time monitoring and early warning system for electromechanical system faults based on digital twins proposed in this invention. Detailed Implementation
[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0020] refer to Figure 1 , Figure 2 and Figure 3 A real-time monitoring and early warning system for electromechanical system faults based on digital twins, including: The data processing module is used to collect operating data of the electromechanical system and perform alignment processing to generate standard electromechanical status data; The twin synchronization module is used to construct a digital twin based on standard electromechanical status data and component connection relationships, synchronize the current state, and generate twin synchronization status data. The perturbation response module is used to apply local virtual perturbations in opposite directions to the component under analysis based on twin synchronization state data, collect the responses of related components, and generate dual perturbation response fringes. The curvature delay mapping module is used to extract response gain, response curvature, response delay and recovery features based on dual perturbation response ripples, and generate bidirectional curvature delay maps; The response analysis module is used to input the bidirectional curvature delay map into the improved GDN model, which has a dual perturbation response coding layer, a curvature delay map learning layer and a mirror neighbor split readout layer, to generate dynamic response features and mirror neighbor split data. The anomaly screening module is used to identify changes in local response relationships based on dynamic response characteristics and generate candidate anomaly components. The degradation confirmation module is used to generate component degradation status data based on the neighbor set and response sorting under opposite perturbation of candidate abnormal components and mirror neighbor order split data. The fault early warning module is used to determine the duration and scope of degradation based on component degradation status data and continuous cycle dynamic response characteristics, and to generate real-time fault monitoring and early warning results.
[0021] In this embodiment, the electromechanical system operation data includes electromechanical operation status data, operation control condition data, electromechanical component configuration data, and historical operation monitoring data.
[0022] In this embodiment, generating standard electromechanical condition data includes: Read electromechanical operating status data, operating control condition data, electromechanical component configuration data, and historical operating monitoring data, and establish data association relationships according to equipment identification, component identification, and data acquisition time; Based on data association, the operating status, control status and operating conditions of each electromechanical component are extracted and arranged according to the time of collection to form a component status sequence. The component status sequence is processed by filtering out abnormal records, filling in missing records, aligning sampling periods, and unifying units to generate preprocessed electromechanical status data. By associating preprocessed electromechanical status data with electromechanical component configuration data according to equipment and component identifiers, a correspondence between component status, component connection, and operating conditions is established, and standard electromechanical status data is generated.
[0023] In this embodiment, generating twin synchronization state data includes: Read component status, control status, operating conditions, component configuration parameters, and component connection relationships from standard electromechanical condition data, and establish a mapping relationship between physical components and twin components according to component identifiers; Establish twin component nodes based on component configuration parameters, connect twin component nodes according to component connection relationships, write control status and operating conditions into the corresponding twin component nodes, and construct a digital twin; Extract normal operation records that match the current operating conditions from historical operation monitoring data, calculate the state deviation between the physical component state and the twin component state, adjust the corresponding twin parameters according to the parameter value range and parameter state response relationship until the state deviation enters the state fluctuation range formed by the normal operation records, and generate twin synchronized state data.
[0024] In this embodiment, generating the dual perturbation response fringe includes: Read the twin synchronization status data, extract adjustable twin parameters, parameter value ranges and component connection relationships from the electromechanical component configuration data, and determine the component to be analyzed according to the component identifier; The twin synchronization state of the component to be analyzed is replicated to establish a first-direction perturbation state, a second-direction perturbation state, and an unperturbed state. The operating conditions, control states, and parameters of the non-analyzed component are kept consistent. Local virtual perturbations with consistent amplitude and opposite directions are applied to the same adjustable twin parameter, wherein: The twin synchronization state of the component to be analyzed is copied to establish a perturbation state in the first direction, a perturbation state in the second direction, and an unperturbed state, specifically as follows: Read the twin synchronization status data at the same sampling time, and extract the node status of the component to be analyzed, the node status of the associated twin component, the control status, the operating condition, the current value of the twin parameter, and the digital twin solution status according to the component identifier; The extracted content is copied into three identical state copies, which are labeled as the first direction perturbation state, the second direction perturbation state, and the unperturbed state, respectively. The starting simulation time of the three state copies is uniformly set to the current sampling time, and the solution step size is uniformly set to the sampling period of the standard electromechanical state data. The external load sequence, control input sequence, environmental input sequence, and component initial state are written into three state copies according to the same sampling index, so that the three state copies have consistent initial state values and operating boundaries before the application of local virtual perturbation; To maintain consistent operating conditions, control states, and parameters of non-analyzed components, apply local virtual perturbations with consistent amplitudes and opposite directions to the same adjustable twin parameter, specifically: Read the lower limit, upper limit and current value of the adjustable twin parameter of the component to be analyzed from the configuration data of the electromechanical components, and calculate the parameter span by the difference between the upper limit and the lower limit of the parameter. Calculate the distance from the current parameter value to the lower limit of the parameter and the distance from the current parameter value to the upper limit of the parameter respectively. Determine the minimum value among 1% of the parameter span, 50% of the distance from the current parameter value to the lower limit of the parameter, and 50% of the distance from the current parameter value to the upper limit of the parameter as the local virtual perturbation amplitude. In the first directional perturbation state, the current parameter value is increased by the local virtual perturbation amplitude; in the second directional perturbation state, the current parameter value is decreased by the local virtual perturbation amplitude; and in the unperturbed state, the current parameter value remains unchanged. Lock all twin parameters of the operating conditions, control states, external inputs and non-analyzed components in the three states, so that the only difference between the first directional perturbation state and the second directional perturbation state is the directional difference of the same adjustable twin parameter. The system operates under three states: a first-direction perturbation state, a second-direction perturbation state, and an unperturbed state. It records the reference response segment before perturbation application, the perturbation-held response segment, and the perturbation-removed recovery response segment at the same sampling period. It then extracts the associated components that generate state responses along the component connection relationships, forming the first-direction response trajectory, the second-direction response trajectory, and the unperturbed response trajectory. The system operates in three states: a first-direction perturbation state, a second-direction perturbation state, and an unperturbed state. Specifically: Three digital twin states are started with the same initial simulation time and the same solution step size, and the first perturbation state, the second perturbation state, and the unperturbed state are synchronously advanced according to the sampling time. For each sampling cycle, the state variables of all twin component nodes are read, the state value corresponding to the first direction perturbation state is written into the first direction record table, the state value corresponding to the second direction perturbation state is written into the second direction record table, and the state value corresponding to the unperturbed state is written into the unperturbed record table. All three record tables are arranged according to component identification, state variable identification, and relative sampling sequence number, so that the same component and the same state variable have corresponding sampling positions in the three operating states; The reference response segment before perturbation application, the perturbation-held response segment, and the perturbation-removed recovery response segment are recorded according to the same sampling period, specifically as follows: Ten consecutive sampling periods were recorded before the application of local virtual perturbation as a reference response segment. The mean and standard deviation of each component and each state variable within the reference response segment were calculated. A local virtual perturbation is applied at the 11th sampling time and maintained continuously for 10 sampling periods. The corresponding recording interval is then defined as the perturbation-holding response segment. At the end of the perturbation hold response segment, the adjustable twin parameters in the first-direction perturbation state and the second-direction perturbation state are restored to their values before the perturbation was applied, and the perturbation removal recovery response segment begins. Each sampling period checks whether each response state variable falls within the range formed by the mean of the corresponding baseline response segment plus or minus 3 times the standard deviation. When all identified associated components fall within the corresponding range for 3 consecutive sampling periods, the recovery record ends. When the recovery record reaches 30 sampling periods and still does not meet the end condition, the recovery record is terminated and an unrecovered mark is written. Extract the associated components that generate state responses along the component connection relationships to form a first-direction response trajectory, a second-direction response trajectory, and an unperturbed response trajectory, specifically as follows: Read the component connection relationship from the electromechanical component configuration data and build an adjacency table according to the component identifier. Use the component to be analyzed as the starting node and search for directly connected components and indirectly connected components layer by layer along the adjacency table. For each component obtained from the search, read the corresponding state variable values at the same sampling time in the first direction record table, the second direction record table, and the unperturbed record table, and calculate the absolute difference between the first direction state value and the unperturbed state value, and the absolute difference between the second direction state value and the unperturbed state value, respectively. The standard deviation of the state variable corresponding to the reference response segment is set as three times the state response threshold. When the state difference in any direction exceeds the state response threshold for two consecutive sampling times, the corresponding component is identified as the associated component. Continue searching for the next layer of components along the component connection relationship from the corresponding sampling time. Stop searching the corresponding branch when there are no new related components in the current level. For all associated components, extract the state values in the first direction, the state values in the second direction, and the unperturbed state values at the sampling time to form the response trajectory in the first direction, the response trajectory in the second direction, and the unperturbed response trajectory. Based on three types of response trajectories—identifier of the component to be analyzed, identifier of the associated component, and identifier of the adjustable twin parameter, and alignment of the sampling time—a baseline response segment, a perturbation-preserved response segment, and a perturbation-removed recovery response segment are combined. A complete response acquisition process is defined as one monitoring cycle, generating a dual perturbation response pattern, wherein: The response trajectories are categorized into three types based on the identifier of the component to be analyzed, the identifier of the associated component, and the identifier of the adjustable twin parameter, and the alignment of the sampling time. Write the identifier of the component to be analyzed, the identifier of the associated component, the identifier of the adjustable twin parameter, the identifier of the state variable, and the relative sampling number for each response record, and define the time when the local virtual perturbation is applied as the relative sampling number 0; Set the sampling sequence number of the baseline response segment to a negative sequence number, and set the sampling sequence numbers of the perturbation hold-up response segment and the perturbation removal recovery response segment to consecutive positive sequences. Establish a joint index according to the identifier of the component to be analyzed, the identifier of the associated component, the identifier of the adjustable twin parameter, the identifier of the state variable, and the relative sampling sequence number. Under the joint index, the first direction response trajectory, the second direction response trajectory, and the undisturbed response trajectory are matched item by item. When a single sampling position is missing a state value, the state values of adjacent sampling positions are read and linear interpolation is performed according to the time distance. When more than two consecutive sampling periods are missing, the corresponding response record is discarded, and the local virtual perturbation of the corresponding component to be analyzed is re-executed; By combining the baseline response segment, the perturbation-preserved response segment, and the perturbation-removed recovery response segment, a complete response acquisition process is defined as a monitoring cycle, generating a dual perturbation response fringe, specifically as follows: The three types of response trajectories are spliced and aligned according to the time sequence of the baseline response segment, the perturbation-preserving response segment, and the perturbation-removal-recovery response segment; The combined results of the associated component identifier and the state variable identifier are used as the row index, the relative sampling sequence number is used as the column index, and the first direction response value, the second direction response value and the unperturbed response value are used as the three response channels. All associated component responses corresponding to the same adjustable twin parameter of the same component to be analyzed are arranged into a three-channel response matrix. A monitoring period is defined as the time from the start of the baseline response segment before the application of the local virtual perturbation to the end of the recovery response segment after the perturbation is removed. The component identifier to be analyzed, the adjustable twin parameter identifier, the local virtual perturbation amplitude, the sampling period, the boundaries of the three response stages, and the three-channel response matrix are associated and encapsulated to generate the dual perturbation response fringe.
[0025] In this embodiment, generating a bidirectional curvature delay map includes: Read the dual perturbation response ripples, and extract the first direction response trajectory, the second direction response trajectory, and the unperturbed response trajectory according to the identifier of the component to be analyzed and the identifier of the associated component; For the same state variable, the state changes of the response trajectories in the first and second directions relative to the unperturbed response trajectory are calculated respectively. The first and second direction response gains are then generated by normalizing the corresponding local virtual perturbation amplitudes. The first and second direction response curvatures are generated based on the changes in the response slopes of the two response trajectories at consecutive sampling times, respectively. Where: For the same state variable, the state changes of the response trajectories in the first and second directions relative to the unperturbed response trajectory are calculated respectively. The first and second direction response gains are then generated by normalizing the corresponding local virtual perturbation amplitudes. Specifically: Read the aligned first-direction response value, second-direction response value, and unperturbed response value according to the identifier of the component to be analyzed, the identifier of the associated component, the identifier of the adjustable twin parameter, the identifier of the state variable, and the relative sampling sequence number, so that the three types of response values correspond one-to-one at each sampling position; For each sampling position, the first direction state change is obtained by subtracting the undisturbed response value at the same sampling position from the first direction response value, and the second direction state change is obtained by subtracting the undisturbed response value at the same sampling position from the second direction response value. The two types of state change values are arranged in order according to their relative sampling numbers to generate the first direction state change sequence and the second direction state change sequence. Read the local virtual perturbation amplitude recorded when generating the dual perturbation response fringe, divide the state change amount in each first direction within the perturbation hold response segment by the local virtual perturbation amplitude, and divide the state change amount in each second direction by the local virtual perturbation amplitude to obtain the normalized response sequence in the first direction and the normalized response sequence in the second direction, respectively. Read the normalized response value with the largest absolute value in the two types of normalized response sequences respectively, take the corresponding absolute value as the first direction response gain and the second direction response gain, and retain the positive or negative sign of the state change corresponding to the largest absolute value as the response direction mark, so that the same state variable forms corresponding response gains in the two perturbation directions. The response curvature in the first direction and the response curvature in the second direction are generated based on the changes in the slope of the response trajectories in the two directions at consecutive sampling times, specifically as follows: Read the sampling period corresponding to the first direction state change sequence, the second direction state change sequence, and the standard electromechanical state data. Calculate the difference in state change for adjacent sampling positions within the perturbation hold response segment, and divide the difference in state change by the sampling period to generate the first direction response slope sequence and the second direction response slope sequence. Starting from the second slope of each response slope sequence, subtract the previous response slope from the current response slope, and then divide by the sampling period to obtain the slope change rate corresponding to each consecutive sampling position. Arrange them according to the sampling order to form the first direction slope change rate sequence and the second direction slope change rate sequence. Calculate the absolute value of each value in the first direction slope change rate sequence and the second direction slope change rate sequence respectively. Determine the maximum absolute value of each direction as the response curvature of the corresponding direction, and record the relative sampling number corresponding to the position of the maximum absolute value to identify the sampling position where the local nonlinear response is most obvious. When the number of valid sampling positions for the perturbation-preserving response segment is less than 3, the corresponding response curvature is not calculated and the response acquisition of the corresponding component to be analyzed is re-executed, so that the change in response slope is formed by at least 3 consecutive sampling positions. A baseline response segment is extracted from the unperturbed response trajectory. The response change boundary is determined by adding or subtracting three standard deviations from the mean of the state variables. The time interval between the application of a perturbation and the first time the response of the associated component crosses the response change boundary is defined as the response delay. The time interval between the removal of the perturbation and the entry of the response of the associated component into the response change boundary at three consecutive sampling times is defined as the recovery characteristic. A baseline response segment is extracted from the unperturbed response trajectory, and the response change boundary is determined by adding or subtracting three times the standard deviation of the state variable mean. Specifically: Read the baseline response segment in the undisturbed response trajectory according to the associated component identifier and state variable identifier, arrange all state variable values of the baseline response segment according to the sampling time, and count the number of valid samples; The mean value of the baseline state is obtained by summing all the state variable values of the baseline response segment and dividing by the number of effective samples. The mean value of the baseline state is obtained by subtracting the mean value of the baseline state from each state variable value and calculating the square of the difference. The square of all the squared differences is summed and divided by the number of effective samples. The square root of the result is used to obtain the standard deviation of the baseline state. The upper boundary of the response change is determined by adding three times the standard deviation of the reference state to the mean of the reference state, and the lower boundary of the response change is determined by subtracting three times the standard deviation of the reference state from the mean of the reference state. The upper boundary and the lower boundary of the response change are used together as the response change boundary of the corresponding state variable of the corresponding associated component. When there are missing sampling locations in the baseline response segment, only the effective state variable values after linear interpolation are used to participate in the calculation of mean and standard deviation. When the missing values are continuously for more than two sampling periods, the boundary calculation of the corresponding state variable is stopped and the corresponding response trajectory is re-acquired. The time interval between the application of a perturbation and the first occurrence of the response exceeding the response change boundary of the associated component is defined as the response delay, specifically: Starting from the moment the local virtual perturbation is applied, the response trajectory in the first direction and the response trajectory in the second direction are read one by one at each sampling position. Each state variable value is compared with the upper boundary and the lower boundary of the response change, respectively. When the value of the state variable is greater than the upper boundary of the response change or less than the lower boundary of the response change for the first time, the corresponding sampling time is recorded. The corresponding sampling time is subtracted from the time when the local virtual perturbation is applied to obtain the response time interval in the corresponding direction, and the response time interval is determined as the response delay in the corresponding direction. When a certain direction does not exceed the response change boundary within the perturbation-holding response segment, the total duration of the perturbation-holding response segment plus one sampling period is determined as the response delay for the corresponding direction, so that the directions that have not triggered obvious responses are arranged after the directions that have generated responses. When neither the first nor the second direction exceeds the response change boundary, cancel the response connection of the current state variable between the component to be analyzed and the corresponding associated component, and do not write the current state variable into the corresponding directed connection; The time interval between perturbation removal and the response of the associated component entering the response change boundary for three consecutive sampling times is defined as the recovery feature, specifically: Starting from the first sampling moment after the removal of the local virtual perturbation, read the corresponding state variable values of the associated components sequentially, and determine whether each state variable value is between the lower boundary and the upper boundary of the response change. When the state variable values of three consecutive sampling times are all within the response change boundary, the first sampling time of the three consecutive sampling times is determined as the recovery start time, and the recovery time is obtained by subtracting the local virtual perturbation removal time from the recovery start time. When the first sampling moment after the perturbation is removed is already within the response change boundary and the subsequent two consecutive sampling moments are still within the response change boundary, the recovery time is recorded as 1 sampling period. When the perturbation removal recovery response segment ends and there are no three consecutive sampling times that enter the response change boundary, the recovery time is recorded as the total duration of the perturbation removal recovery response segment and written into the unrecovered state flag. The recovery time and the recovery state flag together constitute the recovery feature. Using electromechanical components as graph nodes and the response relationships from the component to be analyzed to related components as directed connections, the response gain, response curvature, response delay, and recovery characteristics in each perturbation direction are written into the corresponding directed connections to construct a first-direction curvature delay graph and a second-direction curvature delay graph. These are then aligned according to component identifiers to form a dual curvature delay graph, where: Using electromechanical components as graph nodes, and the response relationships from the component to be analyzed to related components as directed connections, specifically: Read the identifier of the component to be analyzed and all associated component identifiers from the dual perturbation response ripple, perform deduplication for the same component identifier, create one graph node for each unique component identifier, and arrange the graph nodes in ascending order according to the component identifier; Read the state variable values of each component corresponding to the application time of the local virtual perturbation from the unperturbed response trajectory, arrange the state variable values in a fixed order according to the state variable identifier, and write the arrangement result into the corresponding graph node as the node initial feature; Using the graph node corresponding to the component to be analyzed as the starting node and the graph node corresponding to the associated component as the ending node, a directed connection from the starting node to the ending node is established when at least one state variable in the first or second direction generates a valid response. When there are multiple state variables between the same component to be analyzed and the same associated component, only one directed connection is retained. Different state variables are written into the segment feature position of the same directed connection according to the state variable identifier order to avoid duplicate directed connections between the same node pair. The response gain, response curvature, response delay, and recovery characteristics of each perturbation direction are written into the corresponding directed connection, specifically: Read the response gain, response curvature, response delay, and recovery characteristics corresponding to the same directed connection according to the fixed order of the state variable identifiers. For each state variable, arrange the first direction response gain, first direction response curvature, first direction response delay, and first direction recovery characteristics in sequence to generate the first direction connection feature segment. Arrange the second-direction response gain, second-direction response curvature, second-direction response delay, and second-direction recovery features in the order of the same state variables to generate a second-direction connection feature segment, so that the same position in the two directions represents the same type of response feature of the same associated component and the same state variable. When a state variable forms an effective response only in one direction, the response gain and response curvature in the direction where no effective response is formed are recorded as 0, the response delay is recorded as the total duration of the perturbation-held response segment plus 1 sampling period, the recovery time is recorded as 0, and the corresponding response state is marked as not triggered. Write the first direction connection feature segment into the first direction directed connection, write the second direction connection feature segment into the second direction directed connection, and associate it with the identifier of the component to be analyzed, the identifier of the associated component, the identifier of the adjustable twin parameter, and the identifier of the monitoring cycle. Construct curvature delay maps in the first and second directions, and align them according to component identifiers to form a dual curvature delay map, specifically as follows: Combine all graph nodes with the directed connections written into the first direction connection feature segment, arrange the directed connections in ascending order according to the start component identifier and the end component identifier, and generate the first direction curvature delay graph. Combine identical graph nodes with directed connections written into the second-direction connection feature segment, using the curvature delay of the first direction. Figure 1 The node arrangement order and directed connection arrangement order are used to generate a second-direction curvature delay map; Establish node-to-node indexes for two curvature delay graphs using component identifiers, and establish directed connection-to-node indexes using a combination of start and end component identifiers. Pair nodes and directed connections with the same index position in both directions one by one. The curvature delay map in the first direction, the curvature delay map in the second direction, the node corresponding index, the directed connection corresponding index, the identifier of the component to be analyzed, the identifier of the adjustable twin parameter, and the identifier of the monitoring cycle are associated and encapsulated to generate a dual curvature delay map.
[0026] In this embodiment, the generation of dynamic response features and mirror neighbor-order split data includes: Based on the GDN model, the original GDN graph structure learning method, which establishes graph relationships based on node embedding similarity, is replaced with a dynamic response relationship learning method based on dual curvature time delay graphs. A dual perturbation response encoding layer is added at the model input, and a mirror neighbor split readout layer is added at the model output to construct an improved GDN model. Specifically, the construction of the improved GDN model is as follows: Based on the existing GDN model's node feature representation, graph attention aggregation, and temporal state prediction structure, the model is divided into nine functional levels according to data processing functions: dual curvature delay graph input, node and connection feature embedding, dual perturbation response encoding, curvature delay graph learning, graph attention aggregation, temporal state prediction, dynamic response feature output, mirror neighbor order split readout, and mirror neighbor order split data output. Among them, the dual perturbation response encoding structure, curvature delay graph learning structure, and mirror neighbor order split readout structure are newly added or replaced structures, while the graph attention aggregation structure and temporal state prediction structure retain the main processing functions of the GDN model. The nine functional levels are used to describe the data processing order of the model and are not equivalent to a nine-layer trainable neural network. In this embodiment, four types of state variables—vibration RMS, current RMS, temperature, and rotational speed—are selected from the electromechanical operating status data as the original features of the nodes. Therefore, the original input dimension of a single graph node is 4-dimensional. For each perturbation direction, response gain, response curvature, response delay, and recovery features are arranged according to the four types of state variables to form a 16-dimensional directed connection feature. The two directions together form a 32-dimensional connection input. The node features and connection features are mapped to a 64-dimensional representation space. The dual perturbation response encoding structure outputs a 64-dimensional common response representation and a 64-dimensional difference response representation. The curvature delay graph learning structure splices the 64-dimensional representations of the starting node, the ending node, the 64-dimensional common response representation, and the 64-dimensional difference response representation to form a 256-dimensional input and maps it to a 128-dimensional dynamic relationship representation. The graph attention aggregation structure outputs a 64-dimensional current response representation. The temporal state prediction structure forms a 640-dimensional temporal input with the 64-dimensional response representations of 10 consecutive monitoring cycles and outputs a 64-dimensional predicted response representation after passing through a 128-dimensional hidden representation. The dynamic response feature output structure sets up a 64-dimensional to 4-dimensional physical response prediction mapping, converting the predicted response representation into predicted response gain, predicted response curvature, predicted response delay, and predicted recovery features. The absolute difference between the four types of actual dynamic response features and the four types of predicted dynamic response features corresponding to the current monitoring period is calculated to form response gain deviation, response curvature deviation, response delay deviation, and recovery feature deviation. The dynamic response feature consists of a 132-dimensional output composed of a 64-dimensional current response representation, a 64-dimensional predicted response representation, and a 4-dimensional dynamic response deviation. The mirror neighbor order split readout structure outputs 4-dimensional data, including the proportion of non-shared neighbors, the proportion of response gain sorting and exchange, the proportion of response delay sorting and exchange, and the proportion of response curvature sorting and exchange. Therefore, the total dimension of the model output corresponding to a single component to be analyzed is 136 dimensions. In this embodiment, node feature embedding, connection feature embedding, dual perturbation response encoding, 256-dimensional to 128-dimensional curvature-delay dynamic relationship mapping, graph attention aggregation, temporal state prediction, and 64-dimensional to 4-dimensional physical response prediction mapping together constitute the trainable network parameters. Mirror neighbor order split readout uses set comparison and sorting comparison to complete deterministic calculations and does not set trainable parameters. According to the above dimensions, the scale of trainable parameters is approximately 151,000. Four trainable connection feature coefficients are set for adjusting response gain, response curvature, response delay, and recovery feature contribution. The four connection feature coefficients are updated synchronously during training and maintain normalization constraints, so that the model parameter scale meets the online inference requirements under continuous monitoring conditions of electromechanical systems. The model training data comes from continuous normal operation cycles marked as normal and confirmed as fault-free in historical operation monitoring data. These cycles are associated with the corresponding operation control status data and the component configuration parameters and connection relationships in the electromechanical component configuration data to form the basic training data. The normal operation cycles are processed by digital twin synchronization, local virtual perturbation and dual curvature time delay mapping to form the model training input. The model adopts a normal sample self-supervised training method. The trainable parameters are jointly updated by dynamic response prediction constraints and mirror neighbor consistency constraints. During the training phase, continuous neighbor consistency constraints are used to participate in backpropagation. The mirror neighbor split readout structure performs deterministic neighbor set comparison and sorting comparison during the inference phase and does not directly participate in gradient update. The training sample organization method, joint loss calculation method and stopping condition are further executed by the model training process. Compared with existing GDN models, the improved GDN model no longer relies solely on node embedding similarity and candidate neighbor selection to establish graph relationships. Instead, it uses a dual curvature delay graph obtained by controlled perturbation of digital twins to define the neighborhood of the real dynamic response. While retaining the GDN graph attention aggregation mechanism, it uses curvature delay connection weights to modulate the trainable graph attention coefficients, so that the transmission of neighborhood information is constrained by both node state relationships and actual dynamic response relationships. Furthermore, it retains the common and differential responses under local perturbations in opposite directions through the dual perturbation response encoding structure, and increases the comparison between the set of associated components and the response ranking structure in opposite directions through the mirrored neighbor order split readout structure. This expands the model's learning object from the statistical correlation between historical state quantities to the dynamic response transmission relationship caused by changes in local parameters of components. This helps to reduce the interference of statistical correlation changes caused by changes in operating conditions and working point migration, and improves the ability to identify early asymmetric dynamic responses corresponding to gap changes, loosening, friction changes, and asymmetric stiffness degradation. The dual curvature delay map is input into the dual perturbation response encoding layer. Based on the node characteristics and directed connectivity characteristics of the paired directional maps of the analyzed component, associated components, and state variables, directional pairing encoding is performed on the response gain, response curvature, response delay, and recovery characteristics to generate directional common response representations and directional difference response representations, where: The dual curvature delay map is input into the dual perturbation response encoding layer, and the node features and directed connection features in the two directional maps are paired according to the analyzed component, associated components, and state variables, specifically as follows: Read the first-direction curvature delay map and the second-direction curvature delay map from the dual curvature delay map, establish a directed connection joint index according to the identifier of the component to be analyzed and the identifier of the associated component, and establish a graph node joint index according to the component identifier; Under the directed connection joint index, read the connection records of the same analyzed component pointing to the same associated component in two directions, and then read the response gain, response curvature, response delay and recovery characteristics in a fixed order according to the state variable identifier, so that the same index position in two directions corresponds to the same type of response characteristics of the same state variable; Under the graph node joint index, read the node features of the same component in two directions, arrange the node state values in a fixed order according to the state variable identifier, and when a direction is missing a corresponding directed connection, retain the index of the corresponding associated component and write the missing direction into the untriggered flag. The mean and standard deviation of various node features and directed connection features are calculated from the model training samples. The current input is standardized according to the corresponding mean and standard deviation. When the standard deviation is less than 0.000001, the corresponding standard deviation is set to 1 to obtain bidirectional node features and bidirectional connection features with consistent dimensions. Directional pairing encoding is performed on response gain, response curvature, response delay, and recovery characteristics to generate directional common response representations and directional difference response representations, specifically: For the response gain, the perturbation direction correction is completed by multiplying the second-direction normalized response gain by negative 1. The common component of the response gain is obtained by averaging the first-direction normalized response gain and the second-direction normalized response gain after direction correction. The absolute value of the difference between the two is taken as the response gain difference component. For response curvature, response delay and recovery characteristics, the average value of the standardized eigenvalue in the first direction and the standardized eigenvalue in the second direction are calculated respectively to obtain the corresponding common components. The absolute value of the difference between the standardized eigenvalues in the two directions is calculated respectively to obtain the corresponding difference components. The common components of response gain, response curvature, response delay, and recovery characteristics are arranged in the order of state variable identification to form a common connection vector. The four types of difference components are arranged in the same order to form a difference connection vector. The node common vector is formed by averaging the corresponding elements of the node features in the two directions, and the node difference vector is formed by taking the absolute value of the difference between the corresponding elements of the node features in the two directions. The node common vector and the connection common vector are concatenated and input into a 64-dimensional linear mapping layer and rectified linear activation is performed to generate a directional common response representation. The node difference vector and the connection difference vector are mapped with the same dimension to generate a directional difference response representation. The common directional response representation and the directional difference response representation are input into the curvature delay map learning layer. The directed connections in the dual curvature delay map define the neighborhood information transmission range. Connection weights are generated based on response gain, response curvature, response delay, and recovery features. Weighted aggregation and temporal prediction of associated component representations are performed to generate the current response representation and the predicted response representation, which are then combined to form the dynamic response feature, where: The directional common response representation and the directional difference response representation are input into the curvature delay map learning layer, specifically: According to the identifier of the component to be analyzed and the identifier of the associated component, read the 64-dimensional common response representation and the 64-dimensional difference response representation of each directed connection, and concatenate the two representations in dimensional order to form a 128-dimensional connection response vector. Read the common vectors of nodes corresponding to the start and end nodes of the directed connection, and concatenate the start node representation, end node representation and 128-dimensional connection response vector in a fixed order, and input them into the connection feature mapping unit of the curvature delay graph learning layer. The connection feature mapping unit is set with a 128-dimensional hidden output. The mapping result is processed by a rectified linear function to generate a dynamic relationship representation corresponding to each directed connection, and the associated component set is saved according to the component identifier to be analyzed. The directed connections in the dual curvature delay graph define the range of neighborhood information transmission. Connection weights are generated based on response gain, response curvature, response delay, and recovery characteristics, specifically as follows: For each component to be analyzed, only the valid directed connections from the corresponding graph node are read. The components corresponding to the valid directed connections are grouped into the current neighborhood. Components that do not have directed connections in the dual curvature delay graph do not participate in the information aggregation of the current component to be analyzed. Read the standardized response gain, response curvature, response delay, and recovery features of each valid directed connection in the order of state variable identifiers, and average the absolute values of all state variables under the same feature type to form a 4-dimensional connection feature vector. An initial coefficient of 0.25 is set for each of the four types of connectivity features. The same value is used for all four initial coefficients to avoid artificial bias towards a single response feature at the beginning of training. During model training, the joint loss is used to update the four coefficients in reverse. After each update, exponential normalization is performed on the four coefficients to keep the sum of the four coefficients at 1. The original connection scores are obtained by multiplying the 4-dimensional connection features by their corresponding coefficients and summing them. Exponential normalization is then performed on all valid directed connection scores issued by the same component to be analyzed to obtain the connection weights corresponding to each associated component, so that the sum of all connection weights corresponding to the same component to be analyzed is 1. Perform weighted aggregation and time-series prediction of associated component representations to generate current response representations and predicted response representations, and combine them to form dynamic response features, specifically: For each component to be analyzed, read all the 64-dimensional common response representations and 64-dimensional difference response representations of the associated components. Concatenate the two representations into a 128-dimensional associated component representation. Multiply each associated component representation by the corresponding connection weight and sum them dimension by dimension to obtain a 128-dimensional neighborhood aggregation vector. The node representation of the component to be analyzed is concatenated with the 128-dimensional neighborhood aggregation vector, input into a 64-dimensional linear mapping layer and rectified linear activation is performed. The output 64-dimensional vector is determined as the current response representation corresponding to the current monitoring period. Read the 64-dimensional response representations of the 10 consecutive monitoring cycles preceding the current monitoring cycle in the order of the monitoring cycle. Concatenate the 10 response representations into a 640-dimensional time-series vector in chronological order. Input the 640-dimensional time-series vector into a 128-dimensional fully connected prediction layer and perform rectified linear activation. Then input it into a 64-dimensional output layer to generate the predicted response representation corresponding to the current monitoring cycle. When the system has accumulated less than 10 monitoring cycles, it saves the already formed response characterization and pauses time series prediction. After accumulating 10 complete monitoring cycles, it performs prediction and associates and encapsulates the current response characterization, the predicted response characterization, the absolute value of the difference between the corresponding dimensions of the two, the response gain in the first direction and the second direction, the response curvature, the response delay, the recovery feature and the connection weight to generate dynamic response features. The dynamic response features are input into the mirror-neighbor split readout layer. Response gain sorting, response delay sorting, and response curvature sorting are performed on the associated components corresponding to the same component under analysis, according to the first and second directions respectively. The sets of associated components in the two directions are aligned, and the proportions of non-shared neighbors, response gain sorting exchange ratios, response delay sorting exchange ratios, and response curvature sorting exchange ratios are calculated to generate mirror-neighbor split data, where: The response gain sorting, response delay sorting, and response curvature sorting are performed on the associated components corresponding to the same component to be analyzed, respectively, according to the first direction and the second direction. Specifically: Read all effective associated components corresponding to the same component to be analyzed in the first direction from the dynamic response characteristics, calculate the average value of the absolute value of the response gain of all effective state variables, the average value of the response delay and the average value of the response curvature of each associated component, and generate three types of sorting values in the first direction. The first-direction associated components are arranged in descending order of average response gain, in ascending order of average response delay, and in descending order of average response curvature. When two associated components have the same ranking value, their order is determined by ascending order of component identifiers. Using the same calculation rules, all valid associated components in the second direction are read and three types of sorting values in the second direction are generated, forming the second direction response gain sorting sequence, response delay sorting sequence, and response curvature sorting sequence, respectively. Each sorted sequence is saved with the identifier of the component to be analyzed, the identifier of the associated component, the sorting position, and the corresponding sorting value, so that the sorting results in both directions can be compared item by item according to the component identifier. Align the sets of associated components in both directions, calculate the proportions of non-shared neighbors, response gain sorting and swapping, response delay sorting and swapping, and response curvature sorting and swapping, and generate mirrored neighbor-order split data, specifically: Read the first direction associated component set and the second direction associated component set, form a non-shared neighbor set by associated components that exist only in a single direction, merge the two direction associated component sets and remove duplicates to form a joint neighbor set, divide the number of non-shared neighbors by the number of joint neighbors to get the proportion of non-shared neighbors, and record the proportion of non-shared neighbors as 0 when the number of joint neighbors is 0. Extract common neighbors that exist in both directions. When there are at least two common neighbors, perform pairwise combinations on the common neighbors. For each pair of common neighbors, read the response gain sorting position in the first direction and the response gain sorting position in the second direction. When the order of the two neighbors is reversed in the two directions, it is counted as 1 response gain sorting swap. Divide the number of response gain sorting swaps by the total number of common neighbor combinations to obtain the response gain sorting swap ratio. The same pairwise combination method is used to compare the response delay sorting and response curvature sorting respectively to obtain the exchange ratio of response delay sorting and response curvature sorting. When the number of common neighbors is less than two, the exchange ratio of the three sorting types is recorded as 0. Based on the identification of the component to be analyzed and the monitoring cycle, the proportion of non-shared neighbors, the proportion of response gain sorting and swapping, the proportion of response delay sorting and swapping, the proportion of response curvature sorting and swapping, the set of non-shared neighbors, and three types of sorting sequences, mirrored neighbor order split data is generated. Historical state records, operating conditions, control states, component configuration parameters, and component connection relationships corresponding to normal operating cycles are extracted from historical operation monitoring data, operation control condition data, and electromechanical component configuration data. Model training samples are generated through twin synchronization, local virtual perturbation, and dual curvature time delay mapping. The improved Graph Bias Network (GDN) model is trained using a joint loss constructed from dynamic response prediction error and mirror neighbor consistency error. The dual curvature time delay map of the current monitoring cycle is input into the trained model, outputting dynamic response features and mirror neighbor split data, where: The training samples for the model were generated through twin synchronization, local virtual perturbation, and dual curvature time delay mapping, specifically as follows: Extract equipment status markers, maintenance conclusions, historical status records and monitoring cycle identifiers from historical operation monitoring data; filter continuous monitoring cycles where the equipment status marker is normal and the maintenance conclusion is fault-free; extract corresponding operating conditions and control statuses from operation control condition data; and extract corresponding component configuration parameters and component connection relationships from electromechanical component configuration data. Based on the equipment identification and monitoring cycle, historical status records, operating conditions, control status, component configuration parameters and component connection relationships are associated, a twin synchronization status is generated for each monitoring cycle. For each component to be analyzed, local virtual perturbations with the same amplitude rules but opposite directions are performed, and response trajectories in two directions and unperturbed response trajectories are collected. Construct a dual curvature delay map corresponding to each normal monitoring cycle according to the same calculation rules for response gain, response curvature, response delay and recovery characteristics. Combine 11 consecutive normal monitoring cycles into a training sequence. Use the first 10 monitoring cycles as time series input and the 11th monitoring cycle as the dynamic response prediction target. All training sequences were arranged in chronological order of monitoring time. The first 80% of the training sequences were used as the training set, and the last 20% of the training sequences were used as the validation set. Only the training set was used to calculate the mean and standard deviation of the features of each node and the directed connection features, and the statistical results were saved as the model's standardized parameters. The improved Graph Bias Network (GDN) model is trained using a joint loss constructed from the dynamic response prediction error and the mirror neighbor consistency error. Specifically: For each training sample, the current response representation of the 11th monitoring period and the predicted response representation obtained from the previous 10 monitoring periods are read. The squared difference between the two is calculated dimension by dimension and the average is obtained to obtain the dynamic response prediction error. For the same component to be analyzed, perform pairwise combinations of related components in two directions. Calculate the difference between the connection weights in the first direction and the difference between the connection weights in the second direction for each group of related components. When the product of the two differences is less than 0, take the absolute value of the product as the ranking inconsistency of the current neighbor combination. When the product of the two differences is not less than 0, record the ranking inconsistency as 0. Average the ranking inconsistencies of all valid neighbor combinations to obtain the mirror neighbor order consistency error. Before training begins, the first 32 training samples are read from the training set in chronological order and a forward computation without updating the network parameters is performed to obtain the average value of the dynamic response prediction error and the average value of the mirror neighbor consistency error. When either average value is less than 0.000001, the corresponding average value is set to 0.000001. The reciprocals of the two average values are calculated, and then the two reciprocals are divided by the sum of the reciprocals. The two values obtained are used as the weights of the dynamic response prediction error and the mirror neighbor consistency error, respectively, so that the two losses have similar contribution scales at the beginning of training. The dynamic response prediction error is multiplied by the corresponding weight, the mirror neighbor consistency error is multiplied by the corresponding weight and summed to form the joint loss. The adaptive moment estimation optimization algorithm is used to perform backpropagation. The learning rate is set to 0.001, the training batch is set to 32 samples, and the maximum number of training rounds is set to 200 rounds. The learning rate of 0.001 is used to limit the parameter change range in a single round, and the training batch of 32 is used to balance gradient stability and computational cost. After each training round, the joint loss is calculated using the validation set. Training is stopped when the decrease in the joint loss relative to the previous round is no more than 0.001 for 10 consecutive rounds. Training is stopped directly when the number of training rounds reaches 200. The network parameters, four types of connection feature coefficients, node feature normalization parameters, and directed connection feature normalization parameters corresponding to the lowest validation joint loss are saved. Input the dual curvature delay map of the current monitoring period into the trained model, and output dynamic response features and mirror neighbor split data, specifically: Read the dual curvature delay map of the current monitoring period, use the mean and standard deviation saved during the training phase to standardize the node features and directed connection features, complete the pairing of two directional features according to the component to be analyzed, the associated component and the state variable, and generate directional common response representation and directional difference response representation; Read the connection feature coefficients and network parameters saved after training, perform connection weight calculation and associated component representation aggregation on the current dual curvature delay graph, generate the current response representation, read the response representations saved for the previous 10 consecutive monitoring periods and generate the predicted response representation. The current response characterization, the predicted response characterization, and the corresponding response deviation are encapsulated as dynamic response features. The proportion of non-shared neighbors and the proportion of sorting exchange of the three types of sorting sequences are calculated according to the current associated component set in two directions and the three types of sorting sequences, and then encapsulated as mirrored neighbor order split data. The dynamic response characteristics and mirror neighbor sequence split data are saved according to component identification and monitoring cycle, so that the anomaly screening stage can read the output of the two types of models corresponding to the same component and the same monitoring cycle.
[0027] In this embodiment, generating candidate abnormal components includes: Read the dynamic response characteristics and mirror adjacent sequence split data, and extract the response gain deviation, response curvature deviation, response delay deviation and recovery characteristic deviation according to the component identification; Normal operating cycles are extracted from historical operation monitoring data. Based on operation control condition data, normal operating cycles matching the current operating conditions are selected. The 95th percentile values of various dynamic response deviations and mirror-image adjacent sequence split data are calculated to determine the local dynamic response judgment boundary. Among these: The normal operating cycle is extracted from historical operation monitoring data, and the normal operating cycle that matches the current operating condition is selected based on the operation control condition data. Specifically: Read equipment identifiers, component identifiers, monitoring cycle identifiers, equipment status markers, maintenance confirmation results, and historical status records from historical operation monitoring data. Retain monitoring cycles where the equipment status marker is normal and the maintenance confirmation result is fault-free, and establish a normal operation cycle index according to the equipment identifier and monitoring cycle identifier. Read the operating mode, load rate, speed setpoint and control mode corresponding to the current monitoring cycle and each normal operating cycle from the operating control condition data. Perform consistency screening on the operating mode and control mode, and only retain the normal operating cycle with the same value as the current monitoring cycle. Calculate the minimum and maximum values in the historical normal operation data for the load rate and speed setpoints respectively. Divide the absolute value of the difference between the current value and the historical value by the difference between the corresponding maximum and minimum values to obtain the load rate distance and speed distance. When the maximum and minimum values are equal, the corresponding distance is recorded as 0. The normal operating cycle with both load rate distance and speed distance not exceeding 0.1 is determined as the normal operating cycle that matches the current operating condition. When the number of matching cycles is less than 30, the normal operating cycles are supplemented in order of increasing sum of load rate distance and speed distance until the number of matching cycles reaches 30. When the total number of historical normal operating cycles is less than 30, all normal operating cycles that meet the conditions of consistent operating mode and control mode are used. Calculate the 95th percentile values of various dynamic response deviations and mirror-neighbor split data to determine the local dynamic response judgment boundary, specifically: Input the state records corresponding to the normal operation cycle obtained by screening into the improved GDN model after training. Extract the current and predicted values of response gain, response curvature, response delay, and recovery feature according to the component identifier. Calculate the absolute value of the difference between the current and predicted values of the same feature to obtain the response gain deviation, response curvature deviation, response delay deviation, and recovery feature deviation. According to the component identifier, read the proportion of non-shared neighbors, the proportion of response gain sorting and switching, the proportion of response delay sorting and switching, and the proportion of response curvature sorting and switching corresponding to the normal operation cycle, and establish four types of mirror neighbor order splitting numerical sequences respectively. For each component, the response gain deviation, response curvature deviation, response delay deviation, recovery feature deviation, and the values of the four types of mirror neighbor splitting are arranged in ascending order. Assuming the number of effective samples is N, the result of multiplying 0.95 by N is rounded up to obtain the quantile position number. The value at the corresponding position is read as the 95th quantile value of the corresponding feature. The 95th percentile values of response gain deviation, response curvature deviation, response delay deviation, recovery feature deviation, non-common neighbor ratio, and 3 types of sorting exchange ratio for each component are saved as the local dynamic response judgment boundaries for the corresponding component under the corresponding operating conditions. Components whose mirror-image adjacent-order split data exceeds the corresponding judgment boundary are formed into a first candidate set. Components whose response gain deviation, response curvature deviation, response delay deviation, and recovery feature deviation exceed at least two of the corresponding judgment boundaries are formed into a second candidate set. The first and second candidate sets are merged and duplicate components are removed to generate candidate anomalous components, wherein: The components whose mirrored adjacent order split data exceeds the corresponding decision boundary are grouped into the first candidate set, specifically: According to the component identifier, read the proportion of non-shared neighbors, the proportion of response gain sorting and exchange, the proportion of response delay sorting and exchange, and the proportion of response curvature sorting and exchange for the current monitoring period, and read the four types of local dynamic response judgment boundaries corresponding to the current operating conditions of the same component. The four types of mirror neighbor order split values are compared with the corresponding decision boundaries one by one. When at least one type of mirror neighbor order split value is greater than the corresponding decision boundary, the corresponding component identifier, the overbound feature name, the current feature value and the corresponding decision boundary are written into the first candidate record. The first candidate record is deduplicated according to the component identifier. When the same component has multiple mirror neighbor order splits that exceed the boundary, all the outbound feature records are retained, and the deduplicated component set is determined as the first candidate set. Components whose response gain deviation, response curvature deviation, response delay deviation, and recovery characteristic deviation exceed the corresponding judgment boundary are grouped into a second candidate set. The first and second candidate sets are merged and duplicate components are removed to generate candidate anomalous components, specifically: According to the component identifier, read the response gain deviation, response curvature deviation, response delay deviation and recovery characteristic deviation corresponding to the current monitoring cycle, and compare them with the four types of local dynamic response judgment boundaries corresponding to the current operating conditions of the same component. When each deviation is greater than the corresponding judgment boundary, the over-limit count is increased by 1. When the out-of-bounds count of the same component is not less than 2, the corresponding component identifier, out-of-bounds count, out-of-bounds feature name, dynamic response deviation value and corresponding judgment boundary are written into the second candidate record, and the corresponding component is added to the second candidate set; Merge the first and second candidate sets, sort them in ascending order according to the component identifier and perform deduplication. When the same component exists in both candidate sets, merge the mirror adjacent order split out-of-bounds record and the dynamic response deviation out-of-bounds record. The deduplicated component identifier, candidate source, outbound feature name, current feature value, and corresponding judgment boundary are associated and encapsulated to generate candidate abnormal components.
[0028] In this embodiment, the generation of component degradation state data includes: Read candidate abnormal components, continuously extract dynamic response features and mirror neighbor sequence split data according to the monitoring cycle, arrange response curvature deviation, response delay deviation, recovery feature deviation and mirror neighbor sequence split data according to component identifier, and generate candidate component degradation sequence; The characteristic changes of adjacent monitoring cycles in the candidate component degradation sequence are calculated. Normal operating cycles are extracted from historical monitoring data, and the corresponding dynamic response characteristics and mirrored adjacent sequence split data are generated by the response analysis module. The 95th percentile value of the normal cycle characteristic changes is calculated, and monitoring cycles that continuously exceed the corresponding 95th percentile value are determined as degradation duration cycles. Wherein: The characteristic changes of the candidate component degradation sequence over adjacent monitoring periods are calculated as follows: Read the response curvature deviation, response delay deviation, recovery feature deviation, non-common neighbor ratio, response gain sorting and swapping ratio, response delay sorting and swapping ratio, and response curvature sorting and swapping ratio corresponding to the continuous monitoring period in the candidate component degradation sequence according to the component identifier. Arrange each type of feature in the order of the monitoring cycle. Starting from the second monitoring cycle, subtract the feature value of the previous monitoring cycle from the feature value of the current monitoring cycle and take the absolute value to generate the response curvature change, response delay change, recovery feature change and four types of mirror neighbor order change respectively. The seven types of feature changes corresponding to the same component in the same adjacent monitoring cycle are arranged in a fixed order and associated with the previous monitoring cycle identifier, the current monitoring cycle identifier, and the component identifier to form a feature change record for adjacent monitoring cycles. The normal operating cycle is extracted from historical operation monitoring data. The corresponding dynamic response characteristics and mirror-image adjacent sequence split data are generated by the response analysis module. The 95th percentile value of the normal cycle characteristic change is calculated, specifically: Read the equipment status markers and maintenance confirmation results from historical operation monitoring data, retain the continuous monitoring cycles in which the equipment status marker is normal and the maintenance confirmation result is fault-free, and arrange the normal operation cycles according to component identification and monitoring time. The normal operation cycle corresponding to the state record is processed according to the same data processing, twin synchronization, local virtual perturbation, curvature time delay mapping and improved GDN model processing rules to generate dynamic response characteristics and mirror neighbor order split data; For adjacent normal operating cycles, the response curvature change, response delay change, recovery feature change, and four types of mirror neighbor sequence change are obtained according to the same calculation rules as the candidate component degradation sequence. Seven types of normal cycle feature change sequences are established according to component identification. Arrange the sequence of normal periodic feature changes in each type of data from smallest to largest. Multiply 0.95 by the number of effective samples and round up to get the result as the quantile position number. Read the value at the corresponding position as the 95th quantile value of the corresponding feature of the corresponding component and save it as the degradation change judgment value. The monitoring period that continuously exceeds the corresponding 95th percentile value is defined as the degradation duration period, specifically: The changes in the seven types of features corresponding to each adjacent monitoring period of the candidate component are read in the order of the monitoring period, and compared with the seven types of degradation change judgment values corresponding to the same component. The number of features exceeding the corresponding degradation change judgment value in each monitoring period is counted. When the changes in at least two types of features exceed the corresponding degradation change judgment value in the same monitoring period, the current monitoring period is marked as a degradation change period. When the number of out-of-bounds features is less than two, the count will be continuously reset to zero. Starting from the first degradation change cycle, the counting continues. When three consecutive monitoring cycles are marked as degradation change cycles, the continuous interval from the beginning of the monitoring cycle to the current monitoring cycle is determined as the degradation duration cycle. If the continuous state continues, the degradation duration cycle is extended cycle by cycle. When a non-degradation change cycle occurs in a continuous interval, the current degradation duration cycle ends, and continuous counting is restarted from the next monitoring cycle that meets the degradation change conditions. By associating candidate abnormal components, degradation duration, dynamic response deviation, and mirror neighbor sequence split data, component degradation status data is generated.
[0029] In this embodiment, generating real-time fault monitoring and early warning results includes: Read component degradation status data, extract component connection relationships from electromechanical component configuration data, and identify components in a continuous degradation state and directly connected components according to component identification; Dynamic response characteristics are continuously extracted according to the monitoring cycle to determine whether directly connected components continuously exhibit dynamic response exceeding limits, and the degradation impact range is determined layer by layer along the component connection relationship, including: Based on the continuous extraction of dynamic response characteristics according to the monitoring cycle, it is determined whether the directly connected components continuously exhibit dynamic response out-of-bounds behavior. Specifically: Read the component connection relationships corresponding to the components in a continuously deteriorating state from the electromechanical component configuration data, form the first layer of connected component set by the components that are directly connected to the continuously deteriorating components, and establish search records according to the component identifiers; According to the monitoring time range of the degradation duration cycle, read the dynamic response characteristics of each component in the first layer of connected components set, extract the response gain deviation, response curvature deviation, response delay deviation and recovery characteristic deviation, and read the local dynamic response judgment boundary under the corresponding operating conditions of the same component. The four types of dynamic response deviations are compared with the corresponding judgment boundaries for each monitoring cycle. When the same component exceeds the corresponding judgment boundary for no less than two types of dynamic response deviations in the same monitoring cycle, the current monitoring cycle is marked as a dynamic response out-of-bounds cycle. When the same directly connected component is marked as a dynamic response out-of-bounds cycle for three consecutive monitoring cycles, the corresponding directly connected component is identified as a component with continuous dynamic response out-of-bounds, and the component identifier, the start out-of-bounds monitoring cycle, the number of consecutive out-of-bounds cycles, and the out-of-bounds characteristics are written. The scope of degradation impact is determined layer by layer along the component connection relationships, specifically as follows: Components in a continuously degraded state are identified as Level 0 nodes, and directly connected components with continuously out-of-bounds dynamic responses are identified as Level 1 affected nodes. The Level 0 nodes and Level 1 affected nodes are then written into the set of visited components. Starting from each first-level affected node, read the component connection relationship in the electromechanical component configuration data, obtain the directly connected components that have not yet been written into the visited component set, form the second-level component set to be inspected, and check the second-level components to be inspected according to the same dynamic response out-of-bounds judgment rule. The second-layer component that meets the dynamic response out-of-bounds condition for three consecutive monitoring cycles is identified as the second-layer influence node and added to the set of visited components. Then, starting from the newly added influence node, the next layer of components to be inspected is obtained. Stop searching the current branch when there are no new continuously dynamic response out-of-bounds components in the current layer; end the layer-by-layer search when all branches stop expanding; summarize the continuously degraded state components and the affected nodes of each layer in hierarchical order to generate the degradation impact range. A Level 1 warning is established when the degradation impact is limited to components in a state of continuous degradation; a Level 2 warning is established when the degradation impact extends to directly connected components; and a Level 3 warning is established when the degradation impact extends to connected components beyond the directly connected components. The system associates component identification, dynamic response deviation, mirror adjacent sequence split data, degradation duration, degradation impact range, and warning level to generate real-time fault monitoring and warning results.
[0030] Example 1: In an electromechanical system test platform consisting of a drive motor, coupling, gearbox, rolling bearings, and load actuator, the system collects four types of electromechanical operating status data: effective vibration value, effective current value, temperature, and speed. It also reads the load rate, control mode, and speed setpoint. The effective vibration value of the high-frequency vibration signal is calculated in 1-second windows, with the state characteristic sampling period set to 1 second. The digital twin calculation step size is consistent with the sampling period. To simulate normal operating condition migration and mask early degradation, the load rate varies between 45% and 85%, and the speed setpoint is adjusted between 1420 and 1500 r / min. The equivalent radial stiffness of the bearing support is gradually reduced by changing the bearing housing preload.
[0031] During 72 hours of continuous monitoring, 259,200 sampling locations were generated based on a 1-second sampling period. The data included 428 missing sampling locations and 312 duplicate records. The system deleted the 312 duplicate records and used linear interpolation of adjacent valid state values for the 428 missing locations, resulting in 259,200 sets of standard electromechanical condition data. The effective vibration value ranged from 1.82 to 3.36 mm / s, the effective current value ranged from 18.6 to 27.9 A, and the bearing temperature ranged from 43.1 to 58.7 °C. When the load rate increased from 55% to 80%, the effective vibration value briefly rose to 3.31–3.36 mm / s, exceeding the traditional fixed threshold of 3.2 mm / s, but no component degradation was found during condition verification.
[0032] The system establishes a digital twin. During a certain monitoring cycle, the effective value of the bearing's physical vibration is 2.31 mm / s, the initial twin prediction value is 2.17 mm / s, and after parameter calibration, it is adjusted to 2.29 mm / s, with the state deviation decreasing from 0.14 mm / s to 0.02 mm / s. The system replicates the same twin state to form a first-direction perturbation state, a second-direction perturbation state, and an unperturbed state. The current value of the bearing support's equivalent radial stiffness is 1.02 × 10⁸ N / m, with an allowable range of 0.8 × 10⁸–1.2 × 10⁸ N / m. The perturbation amplitude is determined to be 4 × 10⁵ N / m based on a 1% parameter span. The first direction is adjusted to 1.024 × 10⁸ N / m, the second direction to 1.016 × 10⁸ N / m, and the remaining parameters remain consistent with the operating conditions.
[0033] The system records the baseline response for 10 sampling periods, the perturbation-held response for 10 sampling periods, and the removal and recovery response for a maximum of 30 sampling periods, forming a dual perturbation response pattern. After standardization using the mean and standard deviation of the normal training samples, the response gain in the first direction of the bearing-to-gearbox connection is 0.62, and in the second direction it is 0.84. The response curvature in the first direction is 0.071, and in the second direction it is 0.126. The monitoring response delays in the two directions are 2s and 1s, respectively, and the recovery times are 4s and 8s, respectively. The system uses electromechanical components as graph nodes and the response relationship between the component to be analyzed and related components as directed connections. It writes four types of response features, constructs two directional curvature delay maps, and aligns them to form a dual curvature delay map.
[0034] The model training data was independently derived from historical operational monitoring data, extracting 6840 continuous monitoring cycles that were normal and confirmed to be fault-free during maintenance. These cycles were then associated with operating conditions, control status, component configuration parameters, and component connection relationships. After processing, 6830 sets of continuous time-series samples were generated. Each set used the first 10 monitoring cycles as input and the 11th monitoring cycle as the prediction target. 5464 sets were used for training, and 1366 sets were used for validation. In one training sample, the 4D inputs for the bearing node were 2.08 mm / s, 21.7 A, 47.6℃, and 1486 r / min. The vibration response gain, response curvature, response delay, and recovery time in the first direction were 0.47, 0.046, 2 s, and 4 s, respectively, and in the second direction, they were 0.49, 0.049, 2 s, and 4 s, respectively. The model was trained using an adaptive moment estimation optimization algorithm with an initial learning rate of 0.001 and a batch size of 32. Training was stopped when the decrease in joint loss did not exceed 0.001 for 10 consecutive validation rounds.
[0035] After inputting the improved GDN model trained for the current monitoring cycle, the bearing response gain deviation is 0.118, response curvature deviation is 0.064, response delay deviation is 1.21s, and recovery feature deviation is 3.17s. Three of these deviations exceed the 95th percentile judgment boundary under normal operating conditions. There is one non-shared neighbor in each of the two directions, with a non-shared neighbor ratio of 0.25. The response gain ranking exchange ratio is 0.33, the response delay ranking exchange ratio is 0.33, and the response curvature ranking exchange ratio is 0.50. The system identifies the bearing as a candidate abnormal component. Since at least two types of degradation feature changes exceed the 95th percentile value of normal changes in four consecutive monitoring cycles, continuous degradation is confirmed.
[0036] The gearbox node, directly connected to the bearing, experienced simultaneous out-of-bounds deviations in both response curvature and recovery characteristics for three consecutive monitoring cycles. The coupling node, however, only exhibited single-cycle fluctuations. The system determined the degradation impact area to be the area directly connected to the bearing and gearbox and issued a level-two warning. After shutdown, radial loading and displacement measurements were used to re-identify the equivalent radial stiffness of the bearing support. The result showed an 8.6% decrease compared to the normal baseline value. At the time of the warning, the effective bearing vibration value was 2.68 mm / s, lower than the traditional fixed threshold of 3.2 mm / s.
[0037] To verify the effectiveness, the same original historical monitoring records, the same training and validation divisions, and the same testing cycles were used to compare the present invention with the traditional GDN method. The test set included 1200 normal monitoring cycles, 360 operating condition switching cycles, and 420 degradation monitoring cycles, of which 180 were mild degradation, 140 were moderate degradation, and 100 were obvious faults. The early degradation identification accuracy of the traditional GDN was 82.1%, while that of the present invention was 93.8%. The recall rate for mild degradation increased from 68.3% to 89.4%, the false alarm rate for operating condition switching decreased from 12.6% to 4.1%, the accuracy for locating abnormal components increased from 84.7% to 95.2%, and the accuracy for judging the degradation impact range increased from 79.5% to 92.6%. Using the first time the traditional fixed state threshold was exceeded as a benchmark, the present invention issued an early warning on average 7.8 monitoring cycles earlier, while the traditional GDN issued an early warning on average 3.4 monitoring cycles earlier. The true values of the tests were determined by pre-setting the degradation components, component connection topology, and shutdown status inspection results. The above results show that the present invention can identify local dynamic relationship degradation before the absolute value of the state exceeds the limit, and improve the ability to identify early degradation.
[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A real-time monitoring and early warning system for electromechanical system faults based on digital twins, characterized in that, include: The data processing module is used to collect operating data of the electromechanical system and perform alignment processing to generate standard electromechanical status data; The twin synchronization module is used to construct a digital twin based on standard electromechanical status data and component connection relationships, synchronize the current state, and generate twin synchronization status data; The perturbation response module is used to apply local virtual perturbations in opposite directions to the component under analysis based on twin synchronization state data, collect the responses of related components, and generate dual perturbation response patterns. The curvature delay mapping module is used to extract response gain, response curvature, response delay and recovery features based on dual perturbation response ripples, and generate bidirectional curvature delay maps. The response analysis module is used to input the bidirectional curvature delay map into the improved GDN model, which has a dual perturbation response coding layer, a curvature delay map learning layer and a mirror neighbor split readout layer, to generate dynamic response features and mirror neighbor split data. The anomaly screening module is used to identify changes in local response relationships based on dynamic response characteristics and generate candidate anomaly components. The degradation confirmation module is used to generate component degradation status data based on the neighbor set and response sorting under opposite perturbation of candidate abnormal components and mirror neighbor order split data. The fault early warning module is used to determine the duration and scope of degradation based on component degradation status data and continuous cycle dynamic response characteristics, and to generate real-time fault monitoring and early warning results.
2. The real-time monitoring and early warning system for electromechanical system faults based on digital twins according to claim 1, characterized in that, The electromechanical system operation data includes electromechanical operation status data, operation control condition data, electromechanical component configuration data, and historical operation monitoring data.
3. The real-time monitoring and early warning system for electromechanical system faults based on digital twins according to claim 1, characterized in that, The generation of standard electromechanical condition data includes: Read electromechanical operating status data, operating control condition data, electromechanical component configuration data, and historical operating monitoring data, and establish data association relationships according to equipment identification, component identification, and data acquisition time; Based on data association relationships, the operating status, control status and operating conditions of each electromechanical component are extracted and arranged according to the time of collection to form a component status sequence; The component status sequence is processed by filtering out abnormal records, filling in missing records, aligning sampling periods, and unifying units to generate preprocessed electromechanical status data. By associating preprocessed electromechanical status data with electromechanical component configuration data according to equipment and component identifiers, a correspondence between component status, component connection, and operating conditions is established, and standard electromechanical status data is generated.
4. The real-time monitoring and early warning system for electromechanical system faults based on digital twins according to claim 1, characterized in that, The generation of twin synchronization state data includes: Read component status, control status, operating conditions, component configuration parameters and component connection relationships from standard electromechanical condition data, and establish a mapping relationship between physical components and twin components according to component identifiers; Establish twin component nodes based on component configuration parameters, connect twin component nodes according to component connection relationships, write control status and operating conditions into the corresponding twin component nodes, and construct a digital twin; Extract normal operation records that match the current operating conditions from historical operation monitoring data, calculate the state deviation between the physical component state and the twin component state, adjust the corresponding twin parameters according to the parameter value range and parameter state response relationship until the state deviation enters the state fluctuation range formed by the normal operation records, and generate twin synchronized state data.
5. The real-time monitoring and early warning system for electromechanical system faults based on digital twins according to claim 1, characterized in that, The generation of the dual perturbation response fringe includes: Read the twin synchronization status data, extract adjustable twin parameters, parameter value ranges and component connection relationships from the electromechanical component configuration data, and determine the component to be analyzed according to the component identifier; The twin synchronization state of the component to be analyzed is copied, and the first direction perturbation state, the second direction perturbation state and the unperturbed state are established. The operating conditions, control states and parameters of the non-component to be analyzed are kept consistent. Local virtual perturbations with the same amplitude and opposite directions are applied to the same adjustable twin parameter. Run the first direction perturbation state, the second direction perturbation state, and the unperturbed state respectively. Record the reference response segment before perturbation is applied, the perturbation hold response segment, and the perturbation removal recovery response segment according to the same sampling period. Extract the associated components that generate state responses along the component connection relationship to form the first direction response trajectory, the second direction response trajectory, and the unperturbed response trajectory. Based on three types of response trajectories—identifier of the component to be analyzed, identifier of the associated component, identifier of the adjustable twin parameter, and alignment of the sampling time—a baseline response segment, a perturbation-preserved response segment, and a perturbation-removed recovery response segment are combined. A complete response acquisition process is defined as a monitoring cycle, generating a dual perturbation response pattern.
6. The real-time monitoring and early warning system for electromechanical system faults based on digital twins according to claim 1, characterized in that, The generation of the bidirectional curvature delay map includes: Read the dual perturbation response ripples, and extract the first direction response trajectory, the second direction response trajectory, and the unperturbed response trajectory according to the identifier of the component to be analyzed and the identifier of the associated component; For the same state variable, calculate the state change of the first direction response trajectory and the second direction response trajectory relative to the unperturbed response trajectory, and generate the first direction response gain and the second direction response gain according to the corresponding local virtual perturbation amplitude normalization. Generate the first direction response curvature and the second direction response curvature according to the response slope change of the two direction response trajectories at continuous sampling time. The baseline response segment is extracted from the unperturbed response trajectory. The response change boundary is determined by adding or subtracting three times the standard deviation of the state variable mean. The time interval from when the perturbation is applied to the first time the response of the associated component exceeds the response change boundary is determined as the response delay. The time interval from when the perturbation is removed to when the response of the associated component enters the response change boundary for three consecutive sampling times is determined as the recovery feature. Using electromechanical components as graph nodes and the response relationship between the component to be analyzed and related components as directed connections, the response gain, response curvature, response delay and recovery characteristics of each perturbation direction are written into the corresponding directed connections to construct the first direction curvature delay graph and the second direction curvature delay graph, and then aligned according to the component identifier to form the dual curvature delay graph.
7. The real-time monitoring and early warning system for electromechanical system faults based on digital twins according to claim 1, characterized in that, The generated dynamic response features and mirror-neighbor split data include: Based on the GDN model, the original GDN graph structure learning method based on node embedding similarity to establish graph relationships is replaced by a dynamic response relationship learning method based on dual curvature time delay graph. A dual perturbation response encoding layer is added at the model input and a mirror neighbor order split readout layer is added at the model output to construct an improved GDN model. The dual curvature delay map is input into the dual perturbation response coding layer. According to the node characteristics and directed connection characteristics of the two directional maps of the component to be analyzed, the associated component and the state variable, directional pairing coding is performed on the response gain, response curvature, response delay and recovery characteristics to generate directional common response characterization and directional difference response characterization. The common directional response representation and the different directional response representation are input into the curvature delay graph learning layer. The directed connection in the dual curvature delay graph limits the range of neighborhood information transmission. Connection weights are generated based on response gain, response curvature, response delay and recovery features. Weighted aggregation of associated component representations and time-series prediction are performed to generate the current response representation and the predicted response representation and combine them to form dynamic response features. The dynamic response features are input into the mirror neighbor order split readout layer. The response gain sorting, response delay sorting, and response curvature sorting are performed on the associated components corresponding to the same component to be analyzed according to the first and second directions, respectively. The sets of associated components in the two directions are aligned, and the proportion of non-shared neighbors, the proportion of response gain sorting exchange, the proportion of response delay sorting exchange, and the proportion of response curvature sorting exchange are calculated to generate mirror neighbor order split data. Historical state records, operating conditions, control states, component configuration parameters, and component connection relationships corresponding to normal operating cycles are extracted from historical operation monitoring data, operation control condition data, and electromechanical component configuration data. Model training samples are generated through twin synchronization, local virtual perturbation, and dual curvature time delay mapping. The improved graph bias network (GDN) model is trained using dynamic response prediction error and mirror neighbor consistency error as joint loss. The dual curvature time delay map of the current monitoring cycle is input into the trained model, and the dynamic response characteristics and mirror neighbor split data are output.
8. The real-time monitoring and early warning system for electromechanical system faults based on digital twins according to claim 1, characterized in that, The generation of candidate abnormal components includes: Read the dynamic response characteristics and mirror adjacent sequence split data, and extract the response gain deviation, response curvature deviation, response delay deviation and recovery characteristic deviation according to the component identification; Normal operating cycles are extracted from historical operation monitoring data. Normal operating cycles that match the current operating conditions are selected based on operation control condition data. The 95th percentile values of various dynamic response deviations and mirror adjacent sequence split data are calculated to determine the local dynamic response judgment boundary. Components whose mirrored adjacent order split data exceeds the corresponding judgment boundary are formed into a first candidate set. Components whose response gain deviation, response curvature deviation, response delay deviation, and recovery feature deviation exceed the corresponding judgment boundary are formed into a second candidate set. The first and second candidate sets are merged and duplicate components are removed to generate candidate abnormal components.
9. The real-time monitoring and early warning system for electromechanical system faults based on digital twins according to claim 1, characterized in that, The generated component degradation status data includes: Read candidate abnormal components, continuously extract dynamic response features and mirror neighbor sequence split data according to the monitoring cycle, arrange response curvature deviation, response delay deviation, recovery feature deviation and mirror neighbor sequence split data according to component identifier, and generate candidate component degradation sequence; Calculate the characteristic change of adjacent monitoring cycles of candidate component degradation sequence, extract normal operation cycle from historical operation monitoring data, generate corresponding dynamic response characteristics and mirror adjacent sequence split data through response analysis module, calculate the 95th percentile value of normal cycle characteristic change, and determine the monitoring cycle that continuously exceeds the corresponding 95th percentile value as the degradation duration cycle. By associating candidate abnormal components, degradation duration, dynamic response deviation, and mirror neighbor sequence split data, component degradation status data is generated.
10. The real-time monitoring and early warning system for electromechanical system faults based on digital twins according to claim 1, characterized in that, The generated real-time fault monitoring and early warning results include: Read component degradation status data, extract component connection relationships from electromechanical component configuration data, and identify components in a continuous degradation state and directly connected components according to component identification; Dynamic response characteristics are continuously extracted according to the monitoring cycle to determine whether the dynamic response of directly connected components continues to exceed the limit, and the degradation impact range is determined layer by layer along the component connection relationship. A Level 1 warning is established when the degradation impact is limited to components in a state of continuous degradation; a Level 2 warning is established when the degradation impact extends to directly connected components; and a Level 3 warning is established when the degradation impact extends to connected components beyond the directly connected components. The system associates component identification, dynamic response deviation, mirror adjacent sequence split data, degradation duration, degradation impact range, and warning level to generate real-time fault monitoring and warning results.
