Multimodal causal diagnosis and early warning method and system for gravity energy storage systems
By using multimodal data fusion and physical constraint-based causal diagnosis methods, the problem of accurate early warning of latent faults in gravity energy storage systems has been solved, achieving efficient and accurate fault monitoring and early warning, and improving the system's operational reliability and safety.
Patent Information
- Application Number
- CN202610866934.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-16
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2046-06-16
AI Technical Summary
Existing fault monitoring technologies for gravity energy storage systems cannot accurately capture early latent faults, such as bearing micro-pitting, leading to system instability and safety hazards. Existing causal inference technologies lack physical constraints and have low accuracy under complex operating conditions.
By synchronously collecting multimodal data, constructing a multimodal causal topology graph and introducing physical constraints, and combining counterfactual reasoning and external common-cause exogenous nodes, the causal model is dynamically adjusted to perform latent fault hierarchical early warning. Furthermore, the self-calibration and online updating of the diagnostic model are achieved through healthy steady-state data screening and sliding window design.
It improves the accuracy and adaptability of fault early warning for gravity energy storage systems under complex operating conditions, enabling timely identification of early hidden faults and enhancing the reliability and safety of system operation.
Smart Images

Figure CN122709953B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of gravity energy storage system technology, and in particular to a multimodal causal diagnosis early warning method and system for gravity energy storage systems. Background Technology
[0002] As a novel physical energy storage technology, gravity energy storage systems achieve energy conversion through a motor-driven heavy-duty transmission chain. During its long-cycle, high-frequency charging and discharging operation, core components such as the mechanical transmission chain, bearings, and energy storage converter are subjected to alternating loads and heavy-duty low-speed operating conditions for extended periods. This operating characteristic makes it highly susceptible to latent faults such as bearing micro-pitting, early wear of tooth surfaces, and insulation aging within the system. If these subtle latent faults cannot be accurately detected in the early stages, they often rapidly evolve into sudden shutdown accidents or even severe structural damage, seriously threatening the operational safety and economic benefits of energy storage power stations.
[0003] Existing fault monitoring in gravity energy storage systems primarily relies on fixed threshold alarms for a single mode, such as simple vibration exceeding limits or overcurrent protection. In recent years, some general industrial sectors have begun to introduce multimodal data fusion and basic data-driven causal inference algorithms, such as the standard PC algorithm, for root cause analysis of faults. However, existing general-purpose PC algorithms and other causal inference technologies are entirely data-driven and lack physical constraints on the strong coupling between the mechanical, electrical, and thermal systems and the large inertia characteristics of gravity energy storage systems. The causal topologies constructed under different operating conditions have low accuracy and are not deeply bound to the causal topology. Therefore, they cannot achieve accurate hierarchical early warning for very early hidden faults such as bearing micro-pitting. Summary of the Invention
[0004] The main objective of this invention is to provide a multimodal causal diagnosis early warning method for gravity energy storage systems, aiming to solve the technical problems in the prior art.
[0005] This invention proposes a multimodal causal diagnosis early warning method for gravity energy storage systems, comprising: Simultaneously collect multimodal raw data, real-time operating conditions, and real-time operating data of the gravity energy storage system, and extract multimodal features adapted to latent fault characterization based on the multimodal raw data; Based on the real-time operating conditions, a set of physical constraints matching the operating conditions is obtained, and an improved PC algorithm is dynamically embedded based on the multimodal features and the set of physical constraints to construct a multimodal causal topology graph. Based on the multimodal causal topology graph, external co-causal exogenous nodes are introduced to construct a structural causal model. Counterfactual reasoning intervention is performed based on the structural causal model to obtain purified causal topology and root cause tracing results. Based on the purification causal topology binding multimodal features and combined with the fault root cause tracing results, the causal strength residual is obtained, and the latent fault hierarchical early warning is performed based on the causal strength residual to obtain the fault early warning result. A preset health steady-state data determination rule for the gravity energy storage system is obtained, and real-time operating data is filtered according to the preset health steady-state data determination rule to obtain health steady-state data in order to construct a sliding window; Based on the sliding window and fault warning results, the diagnostic model is self-calibrated and updated online to obtain an optimized diagnostic model that is adapted to the entire life cycle operation, so as to realize multimodal causal diagnosis and early warning of latent faults in gravity energy storage systems.
[0006] This application also provides a multimodal causal diagnosis early warning system for gravity energy storage systems, including: The acquisition module is used to synchronously acquire multimodal raw data, real-time operating conditions and real-time operating data of the gravity energy storage system, and extract multimodal features adapted to the latent fault characterization based on the multimodal raw data. The embedding module is used to obtain the physical constraint set matching the operating conditions according to the real-time operating conditions, and dynamically embed the improved PC algorithm according to the multimodal features and the physical constraint set to construct a multimodal causal topology graph. The construction module is used to introduce external common cause exogenous nodes based on the multimodal causal topology graph to construct a structural causal model, and to perform counterfactual reasoning intervention operations based on the structural causal model to obtain purified causal topology and failure root cause tracing results; The early warning module is used to obtain the causal intensity residual based on the purification causal topology binding multimodal features and the fault root cause tracing results, and to perform latent fault hierarchical early warning based on the causal intensity residual to obtain the fault early warning result. The filtering module is used to obtain the preset health steady-state data judgment rules of the gravity energy storage system, and to filter the real-time operating data according to the preset health steady-state data judgment rules to obtain health steady-state data in order to construct a sliding window; The update module is used to perform self-calibration and online incremental update of the diagnostic model based on the sliding window and the fault warning results, so as to obtain an optimized diagnostic model that is adapted to the entire life cycle operation, so as to realize multimodal causal diagnosis and early warning of latent faults of gravity energy storage system.
[0007] The beneficial effects of this invention are as follows: By constructing a multimodal causal topology graph based on real-time operating conditions and physical constraints, this invention enhances the accuracy of the causal inference model. Addressing the complexity of such strongly coupled and large-inertia systems, it introduces external common-cause exogenous nodes to construct a structural causal model and performs counterfactual reasoning, allowing dynamic adjustments under various operating conditions. This ensures the accuracy and adaptability of fault warnings. It utilizes causal strength residuals to provide layered warnings for latent faults, effectively identifying early latent faults such as bearing micro-pitting. Through healthy steady-state data and a sliding window design, the self-calibration and incremental update method of the diagnostic model enables the constructed diagnostic model to adapt to changes throughout its entire lifecycle, achieving more efficient and accurate fault monitoring and warning, and improving the operational reliability of gravity energy storage systems. Attached Figure Description
[0008] Figure 1 This is a schematic diagram of a method flow according to an embodiment of the present invention.
[0009] Figure 2 This is a schematic diagram of the system structure according to an embodiment of the present invention.
[0010] Figure 3 This is a schematic diagram of the embedded module structure according to an embodiment of the present invention.
[0011] Figure 4 This is a schematic diagram of the early warning module structure according to an embodiment of the present invention.
[0012] Figure 5 This is a schematic diagram of the screening module structure according to an embodiment of the present invention.
[0013] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0014] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0015] like Figure 1 As shown, this application provides a multimodal causal diagnosis early warning method for gravity energy storage systems, including: S1. Synchronously collect multimodal raw data, real-time operating conditions and real-time operating data of the gravity energy storage system, and extract multimodal features adapted to the latent fault characterization based on the multimodal raw data. The multimodal raw data includes equipment component status signals collected by four types of sensors: vibration, acoustic pattern, winding current and infrared thermal imaging. S2. Obtain the physical constraint set matching the operating conditions based on the real-time operating conditions, and dynamically embed the improved PC algorithm based on the multimodal features and the physical constraint set to construct a multimodal causal topology graph; S3. Based on the multimodal causal topology graph, external common cause exogenous nodes are introduced to construct a structural causal model, and counterfactual reasoning intervention is performed based on the structural causal model to obtain the purified causal topology and the root cause tracing results of the fault. S4. Obtain the causal intensity residual based on the purification causal topology binding multimodal features and the fault root cause tracing results, and perform latent fault hierarchical early warning based on the causal intensity residual to obtain the fault early warning result. S5. Obtain the preset health steady-state data judgment rules of the gravity energy storage system, and filter the real-time operation data according to the preset health steady-state data judgment rules to obtain health steady-state data in order to construct a sliding window; S6. Based on the sliding window and fault warning results, perform self-calibration and online incremental update of the diagnostic model to obtain an optimized diagnostic model that adapts to the entire life cycle operation, so as to realize multimodal causal diagnosis and early warning of latent faults of the gravity energy storage system.
[0016] As described in steps S1-S6 above, the real-time operating condition refers to the working status labels such as full-load lowering of gravity energy storage, no-load rising, and standby loading, and the real-time operating data refers to the quantifiable and measurable values of the unit's overall operation, such as load weight, operating power, vibration amplitude, current distortion rate, infrared temperature, and equipment fault alarm points.
[0017] This invention improves the ability to capture complex fault modes by synchronously collecting multimodal raw data of gravity energy storage systems and extracting multimodal features adapted to latent fault characterization. It can effectively integrate different types of data and use multi-dimensional information for fault identification. Based on the physical constraint set obtained from real-time operating conditions and combined with multimodal features, an improved PC algorithm is dynamically embedded to construct a multimodal causal topology graph. By introducing physical laws as constraints, the effectiveness of causal inference is improved. By physically modeling the complex mechanical-electrical-thermal coupling characteristics of gravity energy storage systems, the causal topology graph can maintain high accuracy under different operating conditions, thus laying a solid foundation for subsequent root cause analysis of faults.
[0018] Based on a multimodal causal topology graph, an external common-cause exogenous node is introduced to construct a structural causal model. Counterfactual reasoning is used for intervention to obtain purified causal topology and root cause tracing results. By constructing a structural causal model that considers both internal and external factors, it is possible to conduct in-depth root cause analysis of faults and formulate effective countermeasures for potential faults. Combining counterfactual reasoning with the causal model achieves a deep understanding of complex fault mechanisms, providing a profound theoretical foundation for subsequent fault detection and early warning. By combining purified causal topology with multimodal features to obtain causal strength residuals, and using these residuals to perform hierarchical early warning for latent faults, the quantification of causal strength provides hierarchical fault identification capabilities. Different levels of fault warning can effectively improve the monitoring system's response to early latent faults. Compared with traditional fixed threshold alarm mechanisms, it has higher flexibility and accuracy, reduces false alarm rates, and improves system reliability and security.
[0019] By acquiring preset health steady-state data judgment rules for the gravity energy storage system and filtering real-time operating data to construct a sliding window, health assessment within the dynamic time domain is ensured. By introducing health steady-state data judgment into the data filtering process, potential abnormal interference in real-time data is resolved, effectively improving the stability and adaptability of the model. This provides reliable basic data for subsequent model self-calibration. By performing diagnostic model self-calibration and online incremental updates based on the sliding window and fault warning results, the model is ensured to adapt to the entire life cycle operation, realizing a dynamic model update mechanism. This enables the system to self-optimize according to actual operating conditions, maintain efficient fault diagnosis capabilities, and make rapid adaptive adjustments to new operating conditions. This not only improves the system's intelligence level but also achieves continuous fault monitoring and early warning capabilities, thereby improving the operation and maintenance efficiency of the gravity energy storage system.
[0020] In one embodiment, step S2, which involves obtaining a set of physical constraints matching the operating conditions based on the real-time operating conditions and dynamically embedding an improved PC algorithm based on the multimodal features and the set of physical constraints to construct a multimodal causal topology graph, includes: S21. Identify the operating condition type of the gravity energy storage system according to the real-time operating conditions, and match the corresponding electromechanical-thermal coupling relationship and energy transfer flow direction according to the operating condition type to obtain the physical constraint set for operating condition matching. The operating condition types include full-load lowering, no-load rising and standby mounting. S22. Obtain a set of multimodal feature nodes with labeled component attributes, acquisition modality, and spatial location based on the multimodal feature mapping, and perform physical pre-pruning based on the set of multimodal feature nodes and physical constraint set to remove candidate causal edges without physical coupling to obtain an initial constraint skeleton graph; S23. Perform conditional independence test of the improved PC algorithm according to the initial constraint skeleton graph, delete candidate edges with statistically independent conditions, obtain the test-retained edges, and make a preliminary judgment on the test-retained edges according to the V structure rule and Meek propagation rule of the improved PC algorithm to obtain the preliminary judgment result of the edge direction. S24. Match the system energy transfer path according to the working condition type, obtain the working condition-specific energy transfer direction constraint, and perform forced calibration on undirected edges and incorrect direction edges according to the energy transfer direction constraint, the verification of retained edges and the preliminary determination results of edge direction, and construct a multimodal causal topology graph.
[0021] As described in steps S21-S24 above, the improved PC algorithm combines the standard PC algorithm skeleton with a two-stage physical constraint hard embedding, including a four-step execution structure: physical pre-pruning, conditional independence testing, preliminary orientation, and physical forced calibration. Physical pre-pruning refers to directly deleting physically unrelated node pairs based on the spatial coupling constraint matrix and the prohibited edge set, thus narrowing the search space. The spatial coupling constraint matrix is constructed by building a basic topology set of components based on the mechanical transmission chain, electrical connection relationship, and heat conduction path of the gravity energy storage system, and obtaining a component-level spatial topology map by combining the component connection relationship that matches the real-time operating conditions. The shortest topological distance between any two feature nodes is obtained from the component-level spatial topology map, and the shortest topological distance is compared with the preset maximum coupling order to obtain the physical coupling judgment result between nodes. A two-dimensional matrix form of spatial coupling constraint matrix is generated based on the physical coupling judgment result. The spatial coupling constraint matrix is used to determine whether to retain candidate causal edges in the physical pre-pruning stage.
[0022] Conditional independence testing refers to starting from the initial constraint skeleton diagram, checking the conditional independence of nodes in ascending order, deleting independent edges, and obtaining the edges to be retained. Preliminary orientation refers to using V-structure rules to identify collision points and Meek rules to propagate directions, obtaining preliminary results for edge direction determination. Physical forced calibration refers to using working condition energy flow constraints to cover undirected edges and correct erroneous direction edges.
[0023] This invention identifies the operating condition type of a gravity energy storage system based on real-time operating conditions and matches the corresponding electromechanical-thermal coupling relationship with the energy transfer flow direction to obtain a set of physical constraints for operating condition matching. By accurately identifying the operating condition type and combining it with specific electromechanical-thermal coupling relationships, it provides physical characteristic-based constraints for fault monitoring under multiple operating conditions, ensuring high-accuracy fault analysis under different operating conditions. It can more comprehensively reflect the potential fault risks existing under various operating conditions and effectively improve the system's ability to identify early latent faults (such as bearing micro-pitting). Based on multimodal feature mapping, it obtains a set of multimodal feature nodes labeled with component attributes, acquisition modes, and spatial locations. Based on these nodes and the physical constraint set, it performs physical pre-pruning to remove candidate causal edges without physical coupling to obtain an initial constraint skeleton graph. Through physical pre-pruning technology, it effectively eliminates causal edges that do not conform to physical laws, ensuring that the retained causal relationships truly reflect the inherent electromechanical-thermal coupling characteristics of the system, improving the accuracy of the causal topology graph, reducing erroneous judgments caused by false relationships, and thus providing a more robust foundation for fault root cause analysis, ensuring accurate location of fault sources.
[0024] Based on the initial constraint skeleton graph, an improved PC algorithm is performed to check the conditional independence, removing candidate edges with statistically independent conditions. Simultaneously, the edges to be retained are preliminarily judged according to the Meek propagation rule and the V-structure rule. By introducing the improved PC algorithm, compared with traditional methods relying on pure data-driven approaches, it can more effectively verify and retain truly existing causal relationships, improving fault analysis accuracy and reducing unnecessary computational overhead. This achieves efficient and accurate causal structure inference, ensuring timely response and accurate early warning of latent faults. The system's energy transfer path is matched according to the operating condition type to obtain condition-specific energy transfer direction constraints, and undirected edges and edges with incorrect directions are forcibly labeled, thereby constructing a multimodal causal topology graph. By introducing constraints on energy transfer paths, it ensures that causal relationships are not only based on statistical analysis but are also subject to real physical mechanisms, improving the credibility and accuracy of the causal topology graph. Its application effect is particularly significant when facing complex systems and early latent fault detection. It solves the shortcomings of existing causal inference techniques in terms of ambiguous coupling relationships and unclear directional labeling, ensuring accurate hierarchical early warning of faults under complex operating conditions.
[0025] In one embodiment, step S3, which involves constructing a structural causal model by introducing external co-causal exogenous nodes based on the multimodal causal topology graph, and performing counterfactual reasoning intervention based on the structural causal model to obtain purified causal topology and root cause tracing results, includes: S31. Extract measured data of external disturbances and measured data of internal features based on the real-time operation data, and construct a set of external common dependent variables based on the measured data of external disturbances, wherein the observed data of external common variables includes power grid offset data, ambient wind speed data and ambient temperature data; S32. Obtain an internal multimodal feature node set by splitting the multimodal causal topology graph, and determine the external common cause exogenous node based on the internal multimodal feature node set and the external common dependent variable set to set a unidirectional topological constraint, limiting the exogenous node to only be able to point unidirectionally to the internal multimodal feature node. S33. Calculate the sample mean and sample standard deviation based on the measured data of the external disturbance, and obtain the probability distribution parameters of the external variable by fitting the normal distribution. S34. Based on the real-time operating data, healthy steady-state data is obtained to fit the structural equations corresponding to each node in the internal multimodal feature node set, wherein the expression of the structural equation is: ; where X k Let a represent the real-time value of the k-th multimodal feature node. k P represents the correlation coefficient of the k-th parent node. a(k) Represents node X k In a causal topology, the set of all parent node indices, X i Let b represent the value of the feature variable of the i-th topological parent node. k Z represents the correlation coefficient of the k-th external disturbance. z Let ε represent the z-th external common dependent variable. k Let represent the k-th Gaussian perturbation term, and n represent the number of multimodal feature nodes; and construct a structural causal model based on the external variable probability distribution parameters, unidirectional topological constraints, and structural equations. S35. Determine the external common cause health benchmark value based on the health homeostasis data, and perform fixed counterfactual intervention on the external common cause variables of the structural causal model based on the external common cause health benchmark value. Solve the structural equation layer by layer iteratively, that is, fix the external common cause variables to the corresponding health benchmark value, keep the structural equation of the fitted internal nodes unchanged, and perform iterative calculation from the outside to the inside according to the causal topology to obtain the counterfactual values of the internal features without external interference. S36. Based on the difference comparison between the measured data of the internal features and the counterfactual values of the internal features, false causal edges formed by external disturbance coupling are eliminated to obtain the purified causal topology. The node residuals and node topological centrality are obtained based on the purified causal topology and weighted calculations are performed to obtain the root cause tracing results of the fault.
[0026] As described in steps S31-S36 above, the step of determining the exogenous nodes of external common causes is to divide the set of external common causes according to the set of internal multimodal feature nodes and the set of external common causes, set the external common causes variables to point only to internal feature nodes in one direction and prohibit internal feature nodes from pointing to external common causes variables in the opposite direction, and define the exogenous nodes of external common causes according to the unidirectional topological constraints; the step of fitting the structural equation is to split the parent node samples and external common cause input samples corresponding to each internal multimodal feature node according to the healthy steady-state data, take the internal feature values as dependent variables and the parent nodes and external variables as independent variables, and use the regression fitting method to solve the undetermined coefficients of the equation to obtain the structural equation corresponding to each internal node; the node residual is obtained by performing a difference operation on the purified causal topology, the measured values of internal features and the counterfactual values, and the node topological centrality is obtained by statistically calculating the in-degree and out-degree values of each node based on the directed adjacency matrix of the purified causal topology.
[0027] This invention achieves comprehensive perception of the environment and internal system state by extracting measured data of external disturbances and internal features and constructing a set of external codependent variables. By integrating multi-dimensional data such as grid offset, ambient wind speed, and ambient temperature, the richness and diversity of the data are enhanced, effectively capturing the multimodal characteristics of the system's operating state. By decomposing and associating internal multimodal feature nodes with the set of external codependent variables, a unidirectional topological constraint is established, ensuring the unidirectional orientation of external codependent nodes to internal feature nodes. This avoids the generation of potential false causal relationships in causal inference, enhances the model's identifiability, and gives the model higher interpretability, effectively reducing the risk of misjudgment and omission. By statistically analyzing the sample mean and standard deviation of external disturbance data and fitting and calculating the probability distribution parameters of external variables, a reliable statistical basis can be provided for subsequent modeling. By fitting a normal distribution, the dynamic characteristics of external codependent variables can be accurately modeled, which helps to understand the degree of impact of external disturbances on the system and promotes the quantitative assessment of the probability of fault occurrence.
[0028] By screening healthy steady-state data and modeling internal nodes based on structural equations, an accurate structural causal model can be constructed while considering physical laws. This allows for in-depth analysis of the relationships between multimodal feature nodes, further enhancing the system's adaptability to causal relationships under different operating conditions. By fixing the external common dependent variables of the structural causal model through counterfactual intervention and combining health benchmark values to iteratively solve the counterfactual values of internal features, spurious causal relationships under external disturbances can be effectively eliminated. The counterfactual reasoning method brings a new perspective to the in-depth analysis of system faults, enabling hierarchical and precise assessment of fault risks even in complex and variable working environments. This improves the response sensitivity and accuracy of the fault early warning system. By comparing the difference between the measured data of internal features and the counterfactual values, spurious causal edges caused by external disturbances are successfully eliminated, forming a purified causal topology. By obtaining node residuals and topological centrality, the relative importance of each node in fault root cause tracing can be effectively quantified, thus providing a highly concentrated and accurate fault detection capability. This provides a practical and effective solution for achieving in-depth fault analysis and rapid response early warning, ensuring the reliability and stability of the gravity energy storage system under different operating conditions.
[0029] In one embodiment, step S4, which involves obtaining causal intensity residuals based on the purified causal topology binding multimodal features and combined with the fault root cause tracing results, and performing latent fault hierarchical early warning based on the causal intensity residuals to obtain fault early warning results, includes: S41. Extract the set of key directed causal edges associated with faults based on the purification causal topology, and obtain the edge-sensitive weight parameters corresponding to various hidden faults based on the set of key directed causal edges associated with faults and the root cause tracing results. S42. Based on the set of key directed causal edges, bind multimodal features to generate a one-to-one association mapping dataset between key directed causal edges and multimodal features; S43. Obtain the baseline causal intensity characteristic parameters of the gravity energy storage system, and calculate the one-sided standardized causal residual based on the baseline causal intensity characteristic parameters and the associated mapping dataset; S44. Obtain the causal intensity residual of a single type of fault based on the unilateral standardized causal residual and the edge-sensitive weight parameter, and compare the causal intensity residual with the preset hierarchical dynamic early warning threshold to obtain the fault early warning result. The formula for calculating the causal intensity residual is as follows: Among them, R c E represents the residual of causality strength of type c. c Let ω represent the set of directed causal edges that are key to fault association. c,d S represents the edge-sensitive weight parameter. d S represents the one-sided causal strength of the association mapping dataset. 0,dσ represents the baseline causal intensity characteristic parameter. 0,d This represents the standard deviation of the initial healthy sample. This indicates that the minimum denominator is a zero constant.
[0030] As described in steps S41-S44 above, the preset graded dynamic early warning threshold is formed by dynamically correcting the basic early warning threshold based on the real-time load weight and operating power parameters of the gravity energy storage system. The causal intensity residual is then compared with the graded dynamic early warning threshold and the preset duration criterion to determine the final latent fault early warning result in layers.
[0031] This invention extracts a set of key directed causal edges by purifying the causal topology, and then obtains edge-sensitive weight parameters by combining the results of fault root cause tracing. This solves the inaccuracy of single-modality fixed threshold alarms in existing gravity energy storage system fault monitoring. By constructing a key causal relationship network, the accuracy and flexibility of fault mode recognition are enhanced, thus providing a reliable basis for the detection of latent faults and improving the stability and response speed of the system under different operating conditions. By binding the set of key directed causal edges with multimodal features to generate a one-to-one corresponding association mapping dataset, the limitations caused by the single data source in existing technologies are overcome. By integrating data features from different modalities, not only is the comprehensiveness of fault detection improved, but also the complex dynamic interaction relationships in the system are effectively captured, providing multi-dimensional evidence for early warning of faults, improving the detection capability of latent faults such as bearing micro-pitting, and reducing the risk of false alarms.
[0032] By acquiring the baseline causal intensity characteristic parameters of the gravity energy storage system and calculating the one-sided standardized causal residuals using the correlation mapping dataset, a quantitative assessment of the system's health status is achieved. This not only enhances the monitoring capability of system performance but also effectively avoids monitoring blind spots caused by low causal topology accuracy under different operating conditions. Standardized causal residual calculation provides a clearer and more scientific basis for monitoring latent faults, thereby improving the accuracy of overall fault early warning. By calculating the causal intensity residuals of single-type faults and combining them with preset hierarchical dynamic early warning thresholds, detailed fault stratification early warning is further realized. This not only improves the timeliness of fault early warning but also solves the uncertainty caused by early latent faults in gravity energy storage systems. Through quantitative comparison of causal intensity residuals and dynamic thresholds, potential fault risks can be quickly identified, and corresponding preventive measures can be taken, fundamentally improving the system's safety and reliability and ensuring stable operation under various operating conditions.
[0033] In one embodiment, step S5, which involves filtering real-time operating data according to the preset healthy steady-state data determination rule to obtain healthy steady-state data in order to construct a sliding window, includes: S51. Extract sub-item monitoring data based on the real-time operation data, wherein the sub-item monitoring data includes vibration amplitude, current harmonic distortion rate, infrared temperature change rate and equipment fault alarm identifier; S52. Retrieve the rated limit parameter and continuous steady-state duration threshold corresponding to each sub-item monitoring data according to the preset health steady-state data judgment rule; S53. Match the corresponding rated limit parameters to each sub-item monitoring data to determine if it exceeds the limit. Combine the fault alarm sign and the preset continuous steady state duration threshold for joint screening to remove data segments that exceed the limit or have fault marks, and obtain the corresponding healthy steady state data. S54. Summarize multiple health steady-state data to generate a temporary health dataset, and extract continuous and uninterrupted time-series segments from the temporary health dataset according to a preset fixed period duration, discarding discrete and fragmented health data to obtain multiple continuous time-series segments. S55. Calculate the number of valid samples corresponding to each continuous time segment, and perform compliance verification in combination with the preset minimum sample threshold. Select a sliding window with a healthy steady state when the number of valid samples is not less than the preset minimum sample threshold. If the number of valid samples is less than the preset minimum sample threshold, the cached data will be re-collected in the next cycle.
[0034] As described in steps S51-S55 above, a sliding window refers to selecting a continuous, uninterrupted time segment with a preset fixed window duration, and the number of valid samples in the continuous time segment is not less than a preset minimum sample threshold.
[0035] This invention, by breaking down real-time operational data and extracting sub-monitoring data, enables independent monitoring of various important operating parameters in gravity energy storage systems. This avoids the limitations of single-mode monitoring and solves the blind spot problem caused by relying solely on fixed threshold alarms in traditional fault monitoring methods. Based on preset healthy steady-state data judgment rules, the invention retrieves the rated limit parameters and continuous steady-state duration thresholds corresponding to each sub-monitoring data, ensuring the rationality of the monitoring data and the scientific validity of the safety thresholds. It not only introduces theoretical basis for data processing but also overcomes the false alarm problem caused by the lack of clear parameter benchmarks in existing technologies. By comparing and analyzing rated limits with actual operational data, the accuracy of fault monitoring is improved, helping to detect potential faults early.
[0036] By performing limit-exceeding judgments on each monitoring item and combining this with fault alarm indicators and preset continuous steady-state duration thresholds for joint filtering, the accuracy of data filtering is improved. This effectively identifies data with true health status and reduces false alarms caused by fault data by removing data that exceeds limits or has fault markers. This significantly enhances the system's accuracy in monitoring hidden faults. By generating temporary health datasets from multiple healthy steady-state data and extracting continuous time-series segments according to preset fixed period durations, stable operating cycles can be clearly defined, avoiding interference from discrete and discontinuous data and ensuring sufficient data for analysis. Sufficient and stable sample data, especially in the context of multimodal data fusion that needs to be considered in gravity energy storage systems, ensures the consistency and reliability of analysis results. By counting the number of valid samples for each continuous time segment and combining it with a preset minimum sample threshold for compliance verification, the validity and reliability of the dataset are ensured. This enables a more accurate description of the health status during the construction of the sliding window. When the number of valid samples is insufficient and data needs to be cached and extended, the system can flexibly adjust the scheme to adapt to changes in different operating conditions, ensuring sufficient support in data-driven causal inference, thereby improving the feasibility and accuracy of monitoring.
[0037] In one embodiment, step S6, which involves performing diagnostic model self-calibration and online incremental updates based on the sliding window and fault warning results to obtain an optimized diagnostic model adapted for the entire lifecycle operation, includes: S61. Obtain the new mean causal intensity of each key directed causal edge based on the continuous temporal segments of the sliding window, and iteratively update the full causal intensity benchmark value based on the new mean causal intensity and the benchmark causal intensity feature parameters to obtain the updated causal intensity benchmark, wherein the calculation formula is: S new =α*S win-avg +(1-α)*S old Among them, S new This represents the updated causal strength benchmark, α represents the preset smoothing coefficient, and S... win-avg S represents the mean of the novel causal strength. old Indicates the baseline causal strength characteristic parameter; S62. Based on the updated causal strength benchmark, synchronously correct the preset hierarchical dynamic early warning threshold, complete the self-calibration of the diagnostic model benchmark parameters, and obtain the calibrated diagnostic model. S63. Extract the fault residual data and corresponding fault classification identifier of the abnormal samples based on the fault warning results, and collect them to obtain the new fault sample dataset. S64. Obtain the existing fault sample dataset formed by continuously archiving the historical fault warning results, and construct an incremental training set based on the newly added fault sample dataset and the existing fault sample dataset. S65. Based on the calibration model baseline parameters of the calibrated diagnostic model, determine the adjustment boundaries of the historical fault correlation parameters, and iteratively adjust the unconstrained parameters according to the residual loss of the incremental training set to obtain the optimized diagnostic model adapted for the entire life cycle. The formula for calculating the incremental iterative optimization loss of the model is as follows: Where L represents the overall loss value of the model, and the optimized diagnostic model is obtained by minimizing the loss. N all G represents the total number of fault samples in the incremental training set. x G represents the measured composite residual of the x-th sample. pre,x This represents the residual output of the current parameter prediction of the diagnostic model after calibration.
[0038] As described in steps S61-S65 above, the process of obtaining the mean causal intensity based on time-series data segments within a sliding window and updating the causal intensity benchmark in this invention enables dynamic adjustment of the system's operating state. This overcomes the problem that traditional fixed-threshold alarm methods cannot adapt to changing operating conditions and large inertia characteristics, significantly improving the accuracy and adaptability of fault alarms. By fitting the causal relationship of the current operating conditions, it can effectively reflect the real-time changes of potential faults. By synchronously correcting the preset hierarchical dynamic early warning threshold to complete the self-calibration of the diagnostic model's benchmark parameters, it ensures a deep fit between the system's fault early warning threshold and the actual operating state, making the multimodal data fusion early warning mechanism more sensitive and flexible, and able to respond promptly to changes in operating conditions, thereby effectively reducing false alarms and missed alarms. By extracting abnormal samples from the fault early warning results and forming a new fault sample dataset, it achieves a comprehensive understanding and generalization ability of potential fault types, enriching the training foundation of the model. Through continuously updated fault datasets, this step ensures that the model training will not stagnate, thereby effectively improving the sensitivity of fault detection.
[0039] By integrating newly added fault sample datasets with existing fault sample datasets to construct an incremental training set, the generalization ability and accuracy of the model in practical application scenarios can be effectively improved. This strengthens the utilization of historical fault data, avoids inaccuracies caused by data sparsity, and enables effective learning and identification of fault modes of the current large inertia system under different operating conditions. By establishing adjustment boundaries for historical fault correlation parameters based on the calibrated diagnostic model and iteratively adjusting unconstrained parameters through the residual loss of the incremental training set, the fault identification model can be dynamically optimized for various state changes. This makes the optimization process rational and adaptive, enabling not only quantitative assessment of the impact of different faults but also providing accurate hierarchical early warnings when faults occur.
[0040] like Figures 2-5 As shown, this application also provides a multimodal causal diagnosis early warning system for gravity energy storage systems, comprising: The acquisition module is used to synchronously acquire multimodal raw data, real-time operating conditions and real-time operating data of the gravity energy storage system, and extract multimodal features adapted to the latent fault characterization based on the multimodal raw data. The embedding module is used to obtain the physical constraint set matching the operating conditions according to the real-time operating conditions, and dynamically embed the improved PC algorithm according to the multimodal features and the physical constraint set to construct a multimodal causal topology graph. The construction module is used to introduce external common cause exogenous nodes based on the multimodal causal topology graph to construct a structural causal model, and to perform counterfactual reasoning intervention operations based on the structural causal model to obtain purified causal topology and failure root cause tracing results; The early warning module is used to obtain the causal intensity residual based on the purification causal topology binding multimodal features and the fault root cause tracing results, and to perform latent fault hierarchical early warning based on the causal intensity residual to obtain the fault early warning result. The filtering module is used to obtain the preset health steady-state data judgment rules of the gravity energy storage system, and to filter the real-time operating data according to the preset health steady-state data judgment rules to obtain health steady-state data in order to construct a sliding window; The update module is used to perform self-calibration and online incremental update of the diagnostic model based on the sliding window and the fault warning results, so as to obtain an optimized diagnostic model that is adapted to the entire life cycle operation, so as to realize multimodal causal diagnosis and early warning of latent faults of gravity energy storage system.
[0041] In one embodiment, the embedded module includes: The matching unit is used to identify the operating condition type of the gravity energy storage system according to the real-time operating conditions, and to match the corresponding electromechanical-thermal coupling relationship and energy transfer flow direction according to the operating condition type to obtain the physical constraint set for operating condition matching. The mapping unit is used to obtain a set of multimodal feature nodes with labeled component attributes, acquisition modality, and spatial location based on the multimodal feature mapping, and to perform physical pre-pruning based on the set of multimodal feature nodes and the set of physical constraints to remove candidate causal edges without physical coupling and obtain an initial constraint skeleton graph. The verification unit is used to perform conditional independence verification of the improved PC algorithm based on the initial constraint skeleton graph, and obtain preliminary results of the verification of retained edges and edge directions; An embedded calibration unit is used to match the system energy transfer path according to the operating condition type, obtain the operating condition-specific energy transfer direction constraint, and perform forced calibration on undirected edges and incorrect direction edges according to the energy transfer direction constraint, the verification and retention edge and the preliminary determination result of edge direction, so as to construct a multimodal causal topology graph.
[0042] In one embodiment, the early warning module includes: The extraction unit is used to extract a set of key directed causal edges related to faults based on the purified causal topology, and to obtain edge-sensitive weight parameters corresponding to various hidden faults based on the set of key directed causal edges related to faults and the root cause tracing results. The generation unit is used to bind multimodal features according to the set of key directed causal edges and generate a one-to-one association mapping dataset between key directed causal edges and multimodal features; The acquisition unit is used to acquire the baseline causal intensity characteristic parameters of the gravity energy storage system, and to acquire the one-sided standardized causal residual based on the baseline causal intensity characteristic parameters and the correlation mapping dataset. The early warning unit is used to obtain the causal intensity residual of a single type of fault based on the unilateral standardized causal residual and the edge-sensitive weight parameter, and to compare the causal intensity residual with the preset hierarchical dynamic early warning threshold to obtain the fault early warning result.
[0043] In one embodiment, the filtering module includes: The splitting unit is used to split and extract sub-item monitoring data based on the real-time operation data, wherein the sub-item monitoring data includes vibration amplitude, current harmonic distortion rate, infrared temperature change rate and equipment fault alarm identifier; The retrieval unit is used to retrieve the rated limit parameter and continuous steady-state duration threshold corresponding to each sub-item monitoring data according to the preset health steady-state data judgment rule; The filtering unit is used to match the corresponding rated limit parameters for each sub-item monitoring data to determine if it exceeds the limit. It also combines the fault alarm icon and the preset continuous steady state duration threshold to filter and remove data segments that exceed the limit or have fault markers, thereby obtaining the corresponding healthy steady state data. The extraction unit is used to aggregate multiple health steady-state data to generate a temporary health dataset, and extract continuous and uninterrupted time-series segments from the temporary health dataset according to a preset fixed period duration to obtain multiple continuous time-series segments. The verification unit is used to count the number of valid samples corresponding to each continuous time segment, and perform compliance verification in combination with a preset minimum sample threshold. When the number of valid samples is not less than the preset minimum sample threshold, a healthy steady-state sliding window is constructed.
[0044] It should be noted that each module and unit in the multimodal causal diagnosis early warning system for gravity energy storage systems corresponds one-to-one with the steps in the multimodal causal diagnosis early warning method for gravity energy storage systems.
[0045] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.
[0046] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.
Claims
1. A multimodal causal diagnosis and early warning method for gravity energy storage systems, characterized in that, include: Simultaneously collect multimodal raw data, real-time operating conditions, and real-time operating data of the gravity energy storage system, and extract multimodal features adapted to latent fault characterization based on the multimodal raw data; Based on the real-time operating conditions, a set of physical constraints matching the operating conditions is obtained, and an improved PC algorithm is dynamically embedded based on the multimodal features and the set of physical constraints to construct a multimodal causal topology graph. Based on the multimodal causal topology graph, external co-causal exogenous nodes are introduced to construct a structural causal model. Counterfactual reasoning intervention is performed based on the structural causal model to obtain purified causal topology and root cause tracing results. Based on the purification causal topology binding multimodal features and combined with the fault root cause tracing results, the causal strength residual is obtained, and the latent fault hierarchical early warning is performed based on the causal strength residual to obtain the fault early warning result. A preset health steady-state data determination rule for the gravity energy storage system is obtained, and real-time operating data is filtered according to the preset health steady-state data determination rule to obtain health steady-state data in order to construct a sliding window; Based on the sliding window and fault warning results, the diagnostic model is self-calibrated and updated online to obtain an optimized diagnostic model that is adapted to the entire life cycle operation, so as to realize multimodal causal diagnosis and early warning of latent faults in gravity energy storage systems.
2. The multimodal causal diagnosis and early warning method for gravity energy storage systems according to claim 1, characterized in that, The step of obtaining the physical constraint set matching the operating conditions based on the real-time operating conditions, and dynamically embedding the improved PC algorithm based on the multimodal features and the physical constraint set to construct a multimodal causal topology graph includes: The operating condition type of the gravity energy storage system is identified based on the real-time operating conditions, and the corresponding electromechanical-thermal coupling relationship and energy transfer flow direction are matched according to the operating condition type to obtain the physical constraint set for operating condition matching. The set of multimodal feature nodes with labeled component attributes, acquisition modality, and spatial location is obtained based on the multimodal feature mapping. Physical pre-pruning is then performed based on the set of multimodal feature nodes and the set of physical constraints to remove candidate causal edges without physical coupling and obtain the initial constraint skeleton graph. Based on the initial constraint skeleton graph, the conditional independence test of the improved PC algorithm is performed to obtain preliminary results of the test-preserved edges and edge directions; The system energy transfer path is matched according to the operating condition type, and the operating condition-specific energy transfer direction constraint is obtained. Based on the energy transfer direction constraint, the verification of retained edges, and the preliminary determination results of edge direction, the undirected edges and edges with incorrect directions are forcibly calibrated to construct a multimodal causal topology graph.
3. The multimodal causal diagnosis and early warning method for gravity energy storage systems according to claim 1, characterized in that, The steps of constructing a structural causal model by introducing external co-causal exogenous nodes based on the multimodal causal topology graph, and performing counterfactual reasoning intervention based on the structural causal model to obtain purified causal topology and root cause tracing results include: Based on the real-time operating data, external disturbance measured data and internal feature measured data are extracted, and a set of external common dependent variables is constructed based on the external disturbance measured data; The internal multimodal feature node set is obtained by splitting the multimodal causal topology graph, and the external common cause exogenous node is determined based on the internal multimodal feature node set and the external common dependent variable set to set a unidirectional topological constraint, which restricts the exogenous node to only be able to point unidirectionally to the internal multimodal feature node. The probability distribution parameters of the external variables are obtained statistically based on the measured data of the external disturbances. Based on the real-time running data, healthy steady-state data is obtained to fit the structural equations corresponding to each node in the internal multimodal feature node set, and a structural causal model is constructed based on the external variable probability distribution parameters, unidirectional topological constraints, and structural equations. Based on the health homeostasis data, external common cause health baseline values are determined, and fixed counterfactual interventions are performed on the external common cause variables of the structural causal model based on the external common cause health baseline values. The internal feature counterfactual values are obtained by iteratively solving the structural equation layer by layer. By comparing the difference between the measured data of the internal features and the counterfactual values of the internal features, false causal edges are eliminated to obtain a purified causal topology. The node residuals and node topological centrality are obtained based on the purified causal topology and then weighted to obtain the root cause tracing result of the fault.
4. The multimodal causal diagnosis early warning method for gravity energy storage systems according to claim 1, characterized in that, The step of obtaining causal intensity residuals based on the purified causal topology binding multimodal features and combined with the fault root cause tracing results, and performing latent fault hierarchical early warning based on the causal intensity residuals to obtain fault early warning results includes: Based on the purification causal topology, extract the set of key directed causal edges associated with faults, and obtain the edge-sensitive weight parameters corresponding to various hidden faults based on the set of key directed causal edges associated with faults and the root cause tracing results. Based on the set of key directed causal edges, bind multimodal features to generate a one-to-one association mapping dataset between key directed causal edges and multimodal features; Obtain the baseline causal intensity characteristic parameters of the gravity energy storage system, and obtain the one-sided standardized causal residual based on the baseline causal intensity characteristic parameters and the correlation mapping dataset; The causal intensity residual of a single type of fault is obtained based on the unilateral standardized causal residual and the edge-sensitive weight parameter. The causal intensity residual is then compared with the preset hierarchical dynamic early warning threshold to obtain the fault early warning result.
5. The multimodal causal diagnosis and early warning method for gravity energy storage systems according to claim 1, characterized in that, The step of filtering real-time operating data according to the preset healthy steady-state data judgment rules to obtain healthy steady-state data in order to construct a sliding window includes: The real-time operating data is broken down and extracted into sub-item monitoring data, which includes vibration amplitude, current harmonic distortion rate, infrared temperature change rate and equipment fault alarm indicators. According to the preset health steady-state data judgment rules, retrieve the rated limit parameters and continuous steady-state duration thresholds corresponding to each sub-item monitoring data; Each sub-item monitoring data is matched with the corresponding rated limit parameter to determine if it exceeds the limit. Combined with the fault alarm icon and the preset continuous steady state duration threshold, the data segments that exceed the limit or have fault markers are filtered out to obtain the corresponding healthy steady state data. A temporary health dataset is generated by aggregating multiple health steady-state data, and continuous, uninterrupted time-series segments are extracted from the temporary health dataset according to a preset fixed period duration to obtain multiple continuous time-series segments. The number of valid samples corresponding to each continuous time segment is counted, and compliance verification is performed in combination with a preset minimum sample threshold. A healthy steady-state sliding window is constructed when the number of valid samples is not less than the preset minimum sample threshold.
6. The multimodal causal diagnosis and early warning method for gravity energy storage systems according to claim 1, characterized in that, The step of performing self-calibration and online incremental updates of the diagnostic model based on the sliding window and fault warning results to obtain an optimized diagnostic model adapted to the entire lifecycle includes: The new mean causal intensity of each key directed causal edge is obtained based on the continuous temporal segments of the sliding window, and the full causal intensity benchmark value is iteratively updated based on the new mean causal intensity and the benchmark causal intensity feature parameters to obtain the updated causal intensity benchmark. Based on the updated causal strength benchmark, the preset hierarchical dynamic early warning threshold is synchronously corrected to complete the self-calibration of the diagnostic model benchmark parameters and obtain the calibrated diagnostic model. Based on the fault warning results, extract the fault residual data and corresponding fault classification labels of the abnormal samples, and collect them to obtain the new fault sample dataset. Obtain the existing fault sample dataset formed by continuously archiving the historical fault warning results, and construct an incremental training set based on the newly added fault sample dataset and the existing fault sample dataset. Based on the calibration model baseline parameters of the calibrated diagnostic model, the adjustment boundaries of the historical fault association parameters are defined, and the unconstrained parameters are iteratively adjusted according to the residual loss of the incremental training set to obtain an optimized diagnostic model that adapts to the entire life cycle.
7. A multimodal causal diagnostic early warning system for gravity energy storage systems, characterized in that, include: The acquisition module is used to synchronously acquire multimodal raw data, real-time operating conditions and real-time operating data of the gravity energy storage system, and extract multimodal features adapted to the latent fault characterization based on the multimodal raw data. The embedding module is used to obtain the physical constraint set matching the operating conditions according to the real-time operating conditions, and dynamically embed the improved PC algorithm according to the multimodal features and the physical constraint set to construct a multimodal causal topology graph. The construction module is used to introduce external common cause exogenous nodes based on the multimodal causal topology graph to construct a structural causal model, and to perform counterfactual reasoning intervention operations based on the structural causal model to obtain purified causal topology and failure root cause tracing results; The early warning module is used to obtain the causal intensity residual based on the purification causal topology binding multimodal features and the fault root cause tracing results, and to perform latent fault hierarchical early warning based on the causal intensity residual to obtain the fault early warning result. The filtering module is used to obtain the preset health steady-state data judgment rules of the gravity energy storage system, and to filter the real-time operating data according to the preset health steady-state data judgment rules to obtain health steady-state data in order to construct a sliding window; The update module is used to perform self-calibration and online incremental update of the diagnostic model based on the sliding window and the fault warning results, so as to obtain an optimized diagnostic model that is adapted to the entire life cycle operation, so as to realize multimodal causal diagnosis and early warning of latent faults of gravity energy storage system.
8. The multimodal causal diagnosis early warning system for gravity energy storage systems according to claim 7, characterized in that, The embedded module includes: The matching unit is used to identify the operating condition type of the gravity energy storage system according to the real-time operating conditions, and to match the corresponding electromechanical-thermal coupling relationship and energy transfer flow direction according to the operating condition type to obtain the physical constraint set for operating condition matching. The mapping unit is used to obtain a set of multimodal feature nodes with labeled component attributes, acquisition modality, and spatial location based on the multimodal feature mapping, and to perform physical pre-pruning based on the set of multimodal feature nodes and the set of physical constraints to remove candidate causal edges without physical coupling and obtain an initial constraint skeleton graph. The verification unit is used to perform conditional independence verification of the improved PC algorithm based on the initial constraint skeleton graph, and obtain preliminary results of the verification of retained edges and edge directions; An embedded calibration unit is used to match the system energy transfer path according to the operating condition type, obtain the operating condition-specific energy transfer direction constraint, and perform forced calibration on undirected edges and incorrect direction edges according to the energy transfer direction constraint, the verification and retention edge and the preliminary determination result of edge direction, so as to construct a multimodal causal topology graph.
9. The multimodal causal diagnosis early warning system for gravity energy storage systems according to claim 7, characterized in that, The early warning module includes: The extraction unit is used to extract a set of key directed causal edges related to faults based on the purified causal topology, and to obtain edge-sensitive weight parameters corresponding to various hidden faults based on the set of key directed causal edges related to faults and the root cause tracing results. The generation unit is used to bind multimodal features according to the set of key directed causal edges and generate a one-to-one association mapping dataset between key directed causal edges and multimodal features; The acquisition unit is used to acquire the baseline causal intensity characteristic parameters of the gravity energy storage system, and to acquire the one-sided standardized causal residual based on the baseline causal intensity characteristic parameters and the correlation mapping dataset. The early warning unit is used to obtain the causal intensity residual of a single type of fault based on the unilateral standardized causal residual and the edge-sensitive weight parameter, and to compare the causal intensity residual with the preset hierarchical dynamic early warning threshold to obtain the fault early warning result.
10. The multimodal causal diagnosis early warning system for gravity energy storage systems according to claim 7, characterized in that, The filtering module includes: The splitting unit is used to split and extract sub-item monitoring data based on the real-time operation data, wherein the sub-item monitoring data includes vibration amplitude, current harmonic distortion rate, infrared temperature change rate and equipment fault alarm identifier; The retrieval unit is used to retrieve the rated limit parameter and continuous steady-state duration threshold corresponding to each sub-item monitoring data according to the preset health steady-state data judgment rule; The filtering unit is used to match the corresponding rated limit parameters for each sub-item monitoring data to determine if it exceeds the limit. It also combines the fault alarm icon and the preset continuous steady state duration threshold to filter and remove data segments that exceed the limit or have fault markers, thereby obtaining the corresponding healthy steady state data. The extraction unit is used to aggregate multiple health steady-state data to generate a temporary health dataset, and extract continuous and uninterrupted time-series segments from the temporary health dataset according to a preset fixed period duration to obtain multiple continuous time-series segments. The verification unit is used to count the number of valid samples corresponding to each continuous time segment, and perform compliance verification in combination with a preset minimum sample threshold. When the number of valid samples is not less than the preset minimum sample threshold, a healthy steady-state sliding window is constructed.
Citation Information
Patent Citations
Intelligent fault diagnosis and maintenance method and system based on dynamic cause-effect graph of multimodal data fusion
CN119760644A
Network fault trend prediction method and system based on artificial intelligence
CN121842005A
