SiC device complex working condition identification method based on knowledge graph

Abstract

The invention discloses a knowledge graph-based SiC device complex working condition identification method, which comprises the following steps of: acquiring multi-source operation data, and preprocessing to obtain standardized multi-source operation data; performing time scale division to form a cross-scale working condition feature set; constructing a multi-scale evolution knowledge graph, and binding cross-scale working condition features; constructing an improved SEIR model, and forming a comprehensive working condition state set; constructing an atlas feedback harmonizing module to obtain a harmonized comprehensive working condition state set; and forming a complex working condition cluster, determining a complex working condition type, and outputting a risk level and early warning information. According to the invention, by constructing the multi-scale evolution knowledge graph and introducing the improved SEIR model and the graph feedback harmonic module, interpretable recognition and early warning of latent, fluctuating and dangerous working conditions of the SiC device in a complex operation environment are realized.

Description

Technical Field

[0001] This invention relates to the field of power electronic device condition monitoring technology, and in particular to a method for identifying complex operating conditions of SiC devices based on knowledge graphs. Background Technology

[0002] With the rapid development of new energy equipment, power electronic devices, and high power density devices, SiC devices are widely used in new energy vehicles, power grid equipment, aerospace, and industrial power conversion due to their high voltage withstand capability, high switching frequency, and high-temperature operating capability. Traditional SiC device condition monitoring and fault diagnosis methods mostly rely on signal analysis on a single time scale or anomaly detection methods based on fixed thresholds. They can usually only identify faults after parameters have obviously exceeded limits or faults have occurred. They are difficult to reflect the continuous process of the evolution of operating conditions from transient anomalies to long-term degradation, and their ability to express the formation mechanism of complex operating conditions is limited, making it difficult to support the early identification and interpretation of latent risks.

[0003] Existing technologies attempt to model device operating states using multi-source data fusion or knowledge graph-based methods. However, most solutions construct knowledge graphs primarily based on static entity relationships, lacking a unified representation of the evolution of operating conditions across different time scales. This makes it difficult to depict the accumulation and transmission of risks at the millisecond, second, minute, and hour levels. Some state-model-based methods typically employ fixed transformation rules or static parameters during state division and evolution, failing to fully integrate cross-scale operating condition characteristics and mechanistic correlation information. This results in insufficient precision in distinguishing between latent, fluctuating, and hazardous states, easily leading to misjudgments or omissions.

[0004] Therefore, how to provide a knowledge graph-based method for identifying complex operating conditions of SiC devices 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 knowledge graph-based method for identifying complex operating conditions of SiC devices. This invention utilizes multi-timescale modeling, knowledge graph representation, and state evolution analysis to identify and assess the risks associated with complex operating conditions formed during the actual operation of SiC devices. By constructing cross-timescale features from operational data, this invention establishes a multi-scale evolutionary knowledge graph that comprehensively describes the evolutionary relationships of operating conditions at the millisecond, second, minute, and hour levels. An improved SEIR model is introduced to dynamically map and adjust the states of operating condition nodes. Simultaneously, a graph feedback reconciliation mechanism is combined to correct and optimize latent risks, fluctuating operating conditions, and recovery states, thereby achieving the identification and risk level determination of complex operating condition clusters. This invention can identify latent risks before operating conditions reach thresholds, possessing advantages such as high accuracy in operating condition identification and timely risk warning.

[0006] A method for identifying complex operating conditions of SiC devices based on knowledge graphs according to an embodiment of the present invention includes: Multi-source operating data of SiC devices are collected during operation, and the multi-source operating data is preprocessed to obtain standardized multi-source operating data. Based on standardized multi-source operational data, time scales are divided to construct cross-scale operating condition features at the millisecond, second, minute, and hour levels, and these features are combined to form a cross-scale operating condition feature set. Based on the cross-scale operating condition feature set, a multi-scale evolutionary knowledge graph is constructed, and the cross-scale operating condition features are bound to the corresponding operating condition nodes respectively; Based on a multi-scale evolutionary knowledge graph, an improved SEIR model is constructed to perform state mapping, marking the working condition nodes as susceptible states, latent states, fluctuating states, dangerous states, and recovery states, and forming a comprehensive set of working condition states. A graph feedback reconciliation module is constructed to perform semantic consistency verification, latent state amplification, and recovery state rollback on the comprehensive operating condition state set, thereby obtaining the reconciled comprehensive operating condition state set. The system identifies connected regions in the multi-scale evolutionary knowledge graph of the harmonized set of integrated working conditions, forming complex working condition clusters. Based on the state distribution, evolution path, and mechanistic relationships corresponding to the complex working condition clusters, the system determines the type of complex working condition and outputs the risk level and early warning information.

[0007] Optionally, the multi-source operating data specifically includes electrical signal data, thermal signal data, structural and environmental parameters, and control behavior data.

[0008] Optionally, the preprocessing of multi-source operating data specifically includes time synchronization, sampling alignment, abnormal data removal, unit unification, normalization conversion, and noise reduction.

[0009] Optionally, the combination forms a cross-scale operating condition feature set, including: The standardized multi-source operational data is divided into time scales of milliseconds, seconds, minutes and hours. A data analysis window and a sliding interval between adjacent windows are set at each time scale. The data segments after each time scale are segmented and processed. For each data segment at each time scale, cross-scale operating condition features are extracted from electrical signal data, thermal signal data, structural and environmental parameters, and control behavior data. The cross-scale operating condition features extracted at each time scale are mapped to the feature sets at the corresponding time scale, and then combined sequentially at the millisecond, second, minute, and hour levels to form a cross-scale operating condition feature set.

[0010] Optionally, the construction of the multi-scale evolutionary knowledge graph includes: Based on the cross-scale operating condition feature set, the millisecond, second, minute and hourly cross-scale operating condition features are respectively mapped to millisecond-level operating condition nodes, second-level operating condition nodes, minute-level operating condition nodes and hourly operating condition nodes, and time scale identifier and sampling time period identifier are recorded for each operating condition node. Based on operating condition nodes, a set of nodes and a set of relationships are constructed for a multi-scale evolutionary knowledge graph. The set of nodes includes operating condition nodes, parameter nodes, mechanism nodes, and equipment status nodes. The set of relationships includes temporal order relationships at the same time scale, correlation relationships between operating condition nodes and parameters, causal relationships between operating condition nodes and mechanisms, spatial correlation relationships between operating condition nodes, and cross-scale correlation relationships between different time scales. Based on four time scales, millisecond-level subgraphs, second-level subgraphs, minute-level subgraphs, and hour-level subgraphs are constructed respectively, and the temporal order relationship and causal relationship between working condition nodes are established in each subgraph; Establish cross-scale relationships between subgraphs at different time scales, including cross-scale transmission relationships from milliseconds to seconds, from seconds to minutes, and from minutes to hours. Establish cross-scale influence relationships for data segments with cross-scale effects to form a multi-scale evolutionary knowledge graph. Version management is implemented for the multi-scale evolutionary knowledge graph, and the node set and relation set are incrementally updated based on the newly added cross-scale working condition features.

[0011] Optionally, forming the corresponding set of operating conditions includes: An improved SEIR model is constructed, which consists of a state management module, a dynamic conversion rate generation module, a feedback control module, and a state evolution calculation module. The state management module is initialized based on the multi-scale evolutionary knowledge graph. The working condition nodes in the multi-scale evolutionary knowledge graph are written into the state management module, and the correspondence between the working condition nodes and the susceptible state, latent state, fluctuating state, dangerous state and recovery state is established to form an initial state set. The cross-scale operating condition features in the cross-scale operating condition feature set are input into the dynamic conversion rate generation module. Based on the cross-scale features and mechanism correlation of each operating condition node, a state conversion rate function map is constructed. The state conversion paths from susceptible state to latent state, from latent state to fluctuating state, from fluctuating state to dangerous state, and from dangerous state to recovery state are recorded in function form, and the corresponding dynamic functionalized state conversion rates are generated respectively. The dynamic functional state transition rate is input into the state evolution calculation module. A time-scale scheduling stack is introduced to schedule state update tasks at the millisecond, second, minute and hour levels. The initial state set is updated in a positive state at different time scales to obtain the updated positive working condition state set. The positive operating condition state set is input into the feedback control module. Based on the control behavior and cooling enhancement characteristics, a conditional gating backoff feedback control path is constructed from the dangerous state to the fluctuating state and from the fluctuating state to the latent state. The state backoff update is then performed to obtain the backoff operating condition state set. The rollback operating condition state set is rewritten into the state management module for state overwriting and replacement storage, forming a comprehensive operating condition state set. The comprehensive operating condition state set is then returned to the dynamic conversion rate generation module to dynamically update the dynamic function-based state conversion rate.

[0012] Optionally, the obtained harmonized integrated operating condition set includes: A graph feedback reconciliation module is constructed, which consists of a state semantic consistency verification unit, a latent risk amplification unit, a recovery and rollback unit, and a graph state synchronization unit. The comprehensive set of working conditions and the multi-scale evolutionary knowledge graph are input into the state semantic consistency verification unit to construct a semantic support strength graph. Based on the causal chain and scale coupling relationship between working condition nodes and mechanism nodes, parameter nodes and historical evolution trajectories in the knowledge graph, the semantic support strength index of each dangerous state node is calculated, and nodes below the index threshold are screened to form a set of working condition nodes to be corrected. In the state semantic consistency verification unit, the semantic support strength graph and the working condition nodes in the set of working condition nodes to be corrected are combined. The state distribution and temporal evolution relationship of adjacent working condition nodes are analyzed in the multi-scale evolutionary knowledge graph. The working condition nodes in the set of working condition nodes to be corrected are downgraded and adjusted to generate a comprehensive working condition state set after semantic correction. The semantically corrected comprehensive working condition state set is input into the latent risk amplification unit. The working condition nodes in the latent state are taken as the center. Combining the time scale evolution relationship and mechanism causal relationship in the multi-scale evolution knowledge graph, adjacent working condition nodes that meet the preset conditions are detected. A risk level cumulative mapping matrix is ​​introduced to map different risk factor combinations as the upgrade path from latent state to dangerous state. The state label of the working condition nodes that meet the latent risk amplification conditions is updated to form the comprehensive working condition state set after latent risk amplification. The comprehensive operating condition state set after latent risk amplification is input into the recovery and rollback unit. The operating condition nodes in the recovery state are selected. Combining the time scale information in the multi-scale evolutionary knowledge graph and standardized multi-source operation data, a stability rollback gating stack is introduced to periodically determine the stability of the recovery state. The node state corresponding to the unstable operating condition node is rolled back from the recovery state to the latent state, thus obtaining the comprehensive operating condition state set after recovery and rollback processing. The integrated working condition state set after recovery and rollback is input into the graph state synchronization unit. Dependency path analysis and conflict detection are performed on the working condition nodes in the multi-scale evolutionary knowledge graph where there are inconsistencies between the original state and the harmonized state. Based on the state conflict, the closed-loop path in the alignment graph is prioritized for state synchronization, and the state attributes of the corresponding working condition nodes in the multi-scale evolutionary knowledge graph are updated to form the harmonized integrated working condition state set.

[0013] Optionally, determining the type of complex working condition and outputting the risk level and early warning information includes: Based on the harmonized comprehensive working condition state set, working condition nodes in latent, fluctuating, and dangerous states are screened to determine the target working condition node set for participating in complex working condition identification. Based on the target working condition node set, the connectivity between target working condition nodes is identified in the multi-scale evolutionary knowledge graph according to the temporal evolution relationship, mechanistic causal relationship and cross-scale correlation relationship between working condition nodes. The set of working condition nodes with continuous correlation relationship is identified as complex working condition cluster. For each complex working condition cluster, the state distribution characteristics of each working condition node within the complex working condition cluster are analyzed, and the distribution of each working condition node in susceptible state, latent state, fluctuating state, dangerous state and recovery state is recorded. The evolution path information corresponding to the complex working condition cluster is obtained by combining the multi-scale evolutionary knowledge graph. The state distribution characteristics and evolution path information corresponding to complex operating condition clusters are correlated with standardized multi-source operation data to obtain correlation characteristics. Based on the correlation characteristics, the complex operating condition clusters are judged to determine the complex operating condition type, the corresponding risk level is determined, and early warning information is generated based on the evolution path information of complex operating condition clusters.

[0014] The beneficial effects of this invention are: This invention proposes a knowledge graph-based method for identifying complex operating conditions of SiC devices. By preprocessing operational data and constructing cross-scale features, it characterizes the changing features of SiC device operating states at multiple time scales, including milliseconds, seconds, minutes, and hours. A multi-scale evolutionary knowledge graph is constructed to reflect the structured knowledge representation of the accumulation and transmission process of complex operating conditions. This invention introduces an improved SEIR model to map operating condition nodes to multiple states, such as susceptible, latent, fluctuating, dangerous, and recoverable, enabling dynamic evolution and adaptive adjustment of operating condition states at different time scales. Simultaneously, through graph feedback harmonization processing, it calibrates the amplification of latent risks, the regression of recoverable states, and state consistency, ultimately identifying complex operating condition clusters and outputting corresponding risk levels and early warning information. This invention overcomes the limitations of traditional threshold detection and single-scale analysis, achieving early identification and interpretable analysis of complex operating conditions with cross-timescale accumulation features and multi-physical mechanism coupling characteristics. It offers the advantages of high timeliness in risk identification and low false positive rate. Attached Figure Description

[0015] 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:

[0016] Figure 1 This is a flowchart of a knowledge graph-based method for identifying complex operating conditions of SiC devices proposed in this invention. Figure 2 This is a schematic diagram of the improved SEIR model for a knowledge graph-based method for identifying complex operating conditions of SiC devices proposed in this invention. Figure 3 This is a schematic diagram of the graph feedback harmonization module of a knowledge graph-based method for identifying complex operating conditions of SiC devices proposed in this invention. Detailed Implementation

[0017] 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.

[0018] refer to Figure 1 , Figure 2 and Figure 3 A knowledge graph-based method for identifying complex operating conditions of SiC devices includes: Multi-source operating data of SiC devices are collected during operation, and the multi-source operating data is preprocessed to obtain standardized multi-source operating data. Based on standardized multi-source operational data, time scales are divided to construct cross-scale operating condition features at the millisecond, second, minute, and hour levels, and these features are combined to form a cross-scale operating condition feature set. Based on the cross-scale operating condition feature set, a multi-scale evolutionary knowledge graph is constructed, and the cross-scale operating condition features are bound to the corresponding operating condition nodes respectively; Based on a multi-scale evolutionary knowledge graph, an improved SEIR model is constructed to perform state mapping, marking the working condition nodes as susceptible states, latent states, fluctuating states, dangerous states, and recovery states, and forming a comprehensive set of working condition states. A graph feedback reconciliation module is constructed to perform semantic consistency verification, latent state amplification, and recovery state rollback on the comprehensive operating condition state set, thereby obtaining the reconciled comprehensive operating condition state set. The system identifies connected regions in the multi-scale evolutionary knowledge graph of the harmonized set of integrated working conditions, forming complex working condition clusters. Based on the state distribution, evolution path, and mechanistic relationships corresponding to the complex working condition clusters, the system determines the type of complex working condition and outputs the risk level and early warning information.

[0019] In this embodiment, the multi-source operating data specifically includes electrical signal data, thermal signal data, structural and environmental parameters, and control behavior data.

[0020] In this embodiment, the preprocessing of multi-source operating data specifically includes time synchronization, sampling alignment, abnormal data removal, unit unification, normalization conversion, and noise reduction.

[0021] In this embodiment, the combination forming a cross-scale operating condition feature set includes: The standardized multi-source operational data is divided into time scales of milliseconds, seconds, minutes and hours. A data analysis window and a sliding interval between adjacent windows are set at each time scale. The data segments after each time scale are segmented and processed. For each data segment at each time scale, cross-scale operating condition features are extracted from electrical signal data, thermal signal data, structural and environmental parameters, and control behavior data. Specifically, the extraction of these cross-scale operating condition features involves: For each timescale data segment, peak values, gradients, and short-term fluctuation characteristics of voltage, current, and rate of change are extracted from electrical signal data; trend changes, fluctuation amplitudes, and thermal resistance drift characteristics of junction temperature, shell temperature, and rate of temperature rise are extracted from thermal signal data; variation ranges and stability characteristics of structural stress, vibration amplitude, and ambient temperature and humidity are extracted from structural and environmental parameters; and coordination deviation characteristics between control command changes, switching frequency adjustments, cooling regulation behaviors, and electrical and thermal responses are extracted from control behavior data to obtain the corresponding cross-scale operating condition characteristics. The cross-scale operating condition features extracted at each time scale are mapped to the feature sets at the corresponding time scale, and then combined sequentially at the millisecond, second, minute, and hour levels to form a cross-scale operating condition feature set.

[0022] In this embodiment, the construction of a multi-scale evolutionary knowledge graph includes: Based on the cross-scale operating condition feature set, millisecond-level, second-level, minute-level, and hour-level cross-scale operating condition features are respectively mapped to millisecond-level operating condition nodes, second-level operating condition nodes, minute-level operating condition nodes, and hour-level operating condition nodes. A time scale identifier and a sampling time period identifier are recorded for each operating condition node. Specifically, mapping the millisecond-level, second-level, minute-level, and hour-level cross-scale operating condition features to millisecond-level, second-level, minute-level, and hour-level operating condition nodes involves: For each millisecond-level, second-level, minute-level, and hour-level operating condition node data segment, the corresponding cross-scale operating condition features are used as node attributes, and the time window to which they belong is marked as the time scale identifier of the node, thus constructing millisecond-level, second-level, minute-level, and hour-level operating condition nodes respectively. Based on operating condition nodes, a multi-scale evolutionary knowledge graph is constructed, comprising a node set and a relationship set. The node set includes operating condition nodes, parameter nodes, mechanism nodes, and equipment state nodes. The relationship set includes temporal order relationships within the same time scale, associations between operating condition nodes and parameters, causal relationships between operating condition nodes and mechanisms, spatial associations between operating condition nodes, and cross-scale associations between different time scales. Parameter nodes represent the basic physical parameters involved in the generation of operating condition nodes, mechanism nodes represent the potential physical mechanism processes of SiC devices, and device status nodes express the operating status of devices. Cross-scale correlations between different time scales refer to the causal and evolutionary relationships between operating condition nodes at different time scales; Based on four time scales, millisecond-level subgraphs, second-level subgraphs, minute-level subgraphs, and hour-level subgraphs are constructed respectively. In each subgraph, the temporal order relationship and causal relationship between work condition nodes are established. Specifically, the construction of millisecond-level, second-level, minute-level, and hour-level subgraphs based on the four time scales is as follows: At each time scale, work condition nodes with the same time scale identifier are classified into corresponding node sets, and time sorting is performed according to the sampling time period identifier of the work condition nodes. Time order relationship is established according to the time sequence relationship. At the same time, based on the mechanism association information corresponding to the cross-scale work condition features in the work condition nodes, causal relationship is established between work condition nodes with causal driving relationship, forming millisecond-level subgraphs, second-level subgraphs, minute-level subgraphs and hour-level subgraphs respectively. Establish cross-scale relationships between subgraphs at different time scales, including cross-scale propagation relationships from milliseconds to seconds, from seconds to minutes, and from minutes to hours. Establish cross-scale influence relationships for data segments with cross-scale effects, forming a multi-scale evolutionary knowledge graph. Specifically, forming the multi-scale evolutionary knowledge graph involves: After constructing subgraphs for each time scale, the transmission and influence relationships between nodes of different time scales are identified based on cross-scale working condition characteristics. Cross-scale association edges are established between nodes, and the four types of time scale subgraphs are merged to form a multi-scale evolutionary knowledge graph describing the evolutionary behavior of working conditions over time and across scales. Version management is implemented for the multi-scale evolutionary knowledge graph, and the node set and relation set are incrementally updated based on the newly added cross-scale working condition features.

[0023] In this embodiment, forming the corresponding set of working conditions includes: An improved SEIR model is constructed, comprising a state management module, a dynamic conversion rate generation module, a feedback control module, and a state evolution calculation module. Specifically, the construction of the improved SEIR model involves: In addition to the four state storage units of the original SEIR model—susceptible state, latent state, dangerous state, and recovery state—a fluctuating state is added to form a five-state management structure. A state management module is constructed to store and manage the states. By introducing a dynamic functionalized state transformation rate based on cross-scale operating condition characteristics on the basis of a fixed transformation path, the original constant parameter set of the SEIR model is replaced to form a dynamic transformation rate generation module. A backtracking feedback path from dangerous state to fluctuating state and from fluctuating state to latent state is added to the original unidirectional path structure to construct a backtrackable state evolution chain, forming a feedback control module. By superimposing a time-scale scheduling stack on the outside of the basic evolution calculation unit, a state evolution calculation module is formed, resulting in an improved SEIR model. The state management module for operating condition nodes is initialized based on a multi-scale evolutionary knowledge graph. Operating condition nodes from the multi-scale evolutionary knowledge graph are written into the state management module, and a correspondence is established between operating condition nodes and susceptible states, latent states, fluctuating states, dangerous states, and recovery states, forming an initial state set. Specifically, forming the initial state set involves: Millisecond, second, minute, and hourly operating condition nodes are read from the multi-scale evolutionary knowledge graph. Based on the cross-scale operating condition characteristics of the operating condition nodes and their correlation with mechanism nodes and parameter nodes, the initial state of each operating condition node is determined. When the operating condition node only shows slight fluctuations and no abnormal trend, it is marked as a susceptible state. When the operating condition node shows early abnormal signs and the mechanism chain is in a weakly activated state, it is marked as a latent state. When the operating condition node shows a moderate and fluctuating abnormal trend, it is marked as a fluctuating state. When the operating condition node has a strong causal relationship with the dangerous mechanism chain and shows obvious abnormal characteristics, it is marked as a dangerous state. When the operating condition node meets the recovery conditions and the abnormal characteristics subside, it is marked as a recovery state, forming an initial state set. The cross-scale operating condition features from the cross-scale operating condition feature set are input into the dynamic transformation rate generation module. Based on the cross-scale features and mechanistic correlation of each operating condition node, a state transformation rate function map is constructed. The state transformation paths from susceptible state to latent state, from latent state to fluctuating state, from fluctuating state to dangerous state, and from dangerous state to recovery state are recorded in functional form, and corresponding dynamic functionalized state transformation rates are generated for each. The construction of the state transition rate function graph is specifically as follows: By reading the cross-scale operating condition features of each operating condition node at the millisecond, second, minute, and hour levels, and combining the mechanistic causal chain and parameter correlation strength in the multi-scale evolutionary knowledge graph, the four state transformation paths—susceptible state to latent state, latent state to fluctuating state, fluctuating state to dangerous state, and dangerous state to recovery state—are established as function mapping relationships. Each transformation path is assigned a function expression that characterizes the degree of influence of cross-scale operating condition feature changes on state transformation. The function expression takes the weighted combination result of the cross-scale operating condition features related to the corresponding state transformation path as input, and the weighted combination result is processed by monotonically compressed mapping to obtain the transformation rate output between 0 and 1. The feature dependency mode of the four types of state transformation paths is recorded with a consistent structure, and a state transformation rate function graph is constructed. The generation of the corresponding dynamic functional state transition rates is specifically as follows: The cross-scale features of the working condition nodes in the cross-scale working condition feature set are sequentially input into the four function mapping paths in the state transition rate function graph. The state transition rate matching the current features of the working condition node is calculated through the function expression of each mapping path. The cross-scale working condition features exhibit different intensities, trends and mechanistic correlations as the time scale changes. The state transition rate function graph is updated in real time according to the feature quantity of the working condition node during the calculation process. The obtained state transition rates are all dynamically variable, and corresponding dynamic functionalized state transition rates are generated respectively. The dynamically function-based state transition rate is input into the state evolution calculation module. A time-scale scheduling stack is introduced to schedule state update tasks at the millisecond, second, minute, and hourly levels. Forward state updates are performed on the initial state set at different time scales to obtain the updated forward operating condition state set, where: A time-scale scheduling stack is a hierarchical time organization structure that manages the state update order of four time scales—millisecond, second, minute, and hour—in a hierarchical manner according to a fixed stack structure. The time scale is used as the base order, and state update operations of different time granularities are layered and pushed into a unified structural container. Each time scale corresponds to an independent stack layer. The process of performing positive state updates on the initial state set at different time scales specifically involves: State evolution is performed layer by layer across four time scales: millisecond, second, minute, and hour. The dynamic functional state transformation rate is compared with the state transition threshold corresponding to the current time scale. When the dynamic functional state transformation rate is greater than the state transition threshold, the state label of the working condition node is updated to the next state in the state chain. Otherwise, the working condition node remains unchanged in its original state. The state chain is a unidirectional state chain from susceptible state to latent state, from latent state to fluctuating state, from fluctuating state to dangerous state, and from dangerous state to recovery state. Finally, a positive working condition state set that comprehensively reflects transient, short-term, medium-term, and long-term changes is formed. The positive operating condition state set is input to the feedback control module. Based on the control behavior and cooling enhancement characteristics, a conditional gating backoff feedback control path is constructed from the dangerous state to the fluctuating state and from the fluctuating state to the latent state. The state backoff update is then performed to obtain the backoff operating condition state set, where: The construction of the conditional gating backoff feedback control path is as follows: Based on the positive operating condition set, operating condition nodes in dangerous and fluctuating states are selected, and the corresponding cross-scale operating condition characteristics, activation degree of mechanism nodes, and control behavior characteristics and cooling enhancement characteristics in standardized multi-source operating data are read. Combined with the activation decay of mechanism nodes, the temperature drop trend in cooling characteristics, and the load relief caused by control behavior, a weighted comprehensive judgment quantity is formed. When the comprehensive judgment quantity exceeds the gate threshold, the backoff path is activated, and conditional gated backoff feedback control paths from dangerous state to fluctuating state and from fluctuating state to latent state are established respectively. The execution status rollback update specifically refers to: When the gating condition is met, the status marker of the corresponding working condition node rolls back to the previous level state along the constructed rollback feedback control path. The status marker of the dangerous state node is updated from dangerous state to fluctuating state, and the status marker of the fluctuating state node is updated from fluctuating state to latent state. When the gating condition is not met, the working condition node maintains its original state without adjustment. After completing the gating judgment and status rollback of all working condition nodes, the set of rolledback working condition states is composed of all the working condition nodes that have undergone rollback processing and the working condition nodes that have not performed rollback. The rollback operating condition status set is rewritten into the status management module for status overwriting and replacement storage, forming a comprehensive operating condition status set.

[0024] In this embodiment, obtaining the harmonized comprehensive operating condition set includes: A graph feedback reconciliation module is constructed, which consists of a state semantic consistency verification unit, a latent risk amplification unit, a recovery and rollback unit, and a graph state synchronization unit. Specifically, the construction of the graph feedback reconciliation module involves: The state semantic consistency verification unit, the latent risk amplification unit, the recovery and rollback unit, and the graph state synchronization unit are sequentially connected to form the graph feedback reconciliation module; The comprehensive set of operating conditions and the multi-scale evolutionary knowledge graph are input into the state semantic consistency verification unit to construct a semantic support strength graph. Based on the causal chain and scale coupling relationship between operating condition nodes and mechanism nodes, parameter nodes, and historical evolution trajectories in the knowledge graph, the semantic support strength index of each dangerous state node is calculated, and nodes below the index threshold are filtered to form a set of operating condition nodes to be corrected, where: The construction of the semantic support strength graph specifically involves: Using the operating condition nodes in the comprehensive operating condition state set as core nodes, the established causal and correlation edges between operating condition nodes and mechanism and parameter nodes, as well as their corresponding historical evolution trajectories at different time scales, are read from the multi-scale evolutionary knowledge graph. The correlation edges are uniformly labeled with strength according to causal relationships, parameter correlation relationships, and time scale coupling relationships, and written into the relation records corresponding to the operating condition nodes as semantic support information. By performing the same processing on each operating condition node and its correlation, a semantic support strength graph is formed. The calculation of the semantic support strength index for each dangerous state node is specifically as follows: For each working condition node marked as dangerous, the strength of the causal relationship of the mechanism, the strength of the parameter association, and the scale coupling strength reflected in the historical evolution trajectory are extracted from the semantic support strength map. The results are then weighted and aggregated to obtain a comprehensive value that reflects the semantic support level of the working condition node. The comprehensive value is then normalized and mapped to a unified index range to serve as the semantic support strength index of the working condition node. In the state semantic consistency verification unit, combining the semantic support strength graph and the work condition nodes in the set of work condition nodes to be corrected, the state distribution and temporal evolution relationship of adjacent work condition nodes are analyzed in the multi-scale evolutionary knowledge graph. State downgrading adjustments are then performed on the work condition nodes in the set of work condition nodes to be corrected, generating a semantically corrected comprehensive work condition state set. Specifically, the state downgrading adjustments are performed on the work condition nodes in the set of work condition nodes to be corrected as follows: The semantic support strength index of the corresponding working condition node in the semantic support strength graph is used as the basis, and a comprehensive judgment is made in combination with the state distribution and temporal evolution relationship of the working condition node adjacent to the corresponding working condition node in the multi-scale evolutionary knowledge graph. When the node to be corrected is in a dangerous state and the semantic support strength index is lower than the index threshold, and the adjacent nodes are mainly distributed in the fluctuating state and the latent state on the same time scale, the state label of the node is downgraded from dangerous state to fluctuating state. When the node to be corrected is in a fluctuating state and the semantic support strength index is lower than the index threshold, and the adjacent nodes are mainly distributed in the latent state on the same time scale, the state label of the node is downgraded from fluctuating state to latent state. By downgrading the semantic support strength and the consistency of neighborhood state evolution, the state label is consistent with the overall semantics and evolution trend reflected in the multi-scale evolutionary knowledge graph, and a semantically corrected comprehensive working condition state set is generated. The semantically corrected comprehensive operating condition state set is input into the latent risk amplification unit. Operating condition nodes in the latent state are used as centers. Combining the temporal-scale evolutionary relationships and causal mechanisms in the multi-scale evolutionary knowledge graph, adjacent operating condition nodes that meet preset conditions are detected. A risk level cumulative mapping matrix is ​​introduced to map different combinations of risk factors as escalation paths from latent to dangerous states. The state labels of operating condition nodes that meet the latent risk amplification conditions are updated, forming the comprehensive operating condition state set after latent risk amplification processing. Where: The preset conditions refer to the following conditions: the latent condition node simultaneously meets the following requirements in the multi-scale evolutionary knowledge graph: the condition node remains in a latent state for at least two adjacent time scales; the condition node has at least one mechanistic node with a mechanistic causal relationship and the corresponding activation level reaches a threshold; and the cross-scale condition feature corresponding to the condition node shows a consistent shift trend during the time scale evolution process and does not belong to a single isolated mutation feature. Risk factors refer to the persistence of abnormal features across time scales, the cumulative trend of abnormal amplitudes, the strength of the correlation with the nodes of the hazard mechanism, and the consistency of the state evolution of adjacent working condition nodes, reflecting the comprehensive changes in the potential risk level of working condition nodes. The risk level cumulative mapping matrix refers to a set of structured mapping relationships that describe the cumulative effect of different risk factors on risk levels during the multi-scale working condition evolution process. It takes the latent state working condition nodes as the mapping objects, and combines and arranges the risk factors corresponding to the cross-scale working condition characteristics, the strength of the causal relationship of the mechanism, and the continuity of the time scale evolution. The latent state working condition nodes form a clear risk accumulation characterization under the joint action of multi-dimensional risk factors, providing a unified mapping basis for the escalation path from latent state to dangerous state. The formation of the comprehensive operating condition set after latent risk amplification is specifically as follows: For latent state work nodes that meet the preset conditions and reach the risk accumulation threshold through the risk level accumulation mapping matrix, the corresponding state label is updated from latent state to dangerous state. For work nodes that do not meet the preset conditions, the original state label remains unchanged. The work nodes that have completed the state update and the work nodes that have not changed the state are uniformly collected to obtain the comprehensive work state set after latent risk amplification processing. The comprehensive operating condition state set after latent risk amplification is input into the recovery and rollback unit. Operating condition nodes in the recovery state are selected. Combining timescale information from the multi-scale evolutionary knowledge graph and standardized multi-source operational data, a stability rollback gating stack is introduced to periodically determine the stability of the recovery state. The node state corresponding to the unstable operating condition node is rolled back from the recovery state to the latent state, resulting in the comprehensive operating condition state set after recovery and rollback processing. The stability rollback gating stack refers to a time-ordered judgment structure that describes the criteria for determining the stability of the recovery state. It uses the time scale as the organizational basis and records and accumulates multiple judgment elements that reflect the reliability of the recovery of the working node in a hierarchical manner according to the time order. It portrays the continuity and consistency of the abnormal decay characteristics, mechanism correlation changes and cross-time scale stability performance of the working node in the recovery state, and provides a structured judgment basis for whether the recovery state has continuous stability. An unstable operating condition node refers to an operating condition node in the recovery state whose corresponding standardized multi-source operating data exhibits abnormal fluctuation characteristics, and whose associated mechanism nodes in the multi-scale evolutionary knowledge graph are in an active state. The operating condition node is thus determined to be an unstable operating condition node. The obtained comprehensive operating condition state set after recovery and rollback processing is specifically as follows: The node state corresponding to the unstable operating condition node is rolled back from the recovered state to the latent state. The remaining operating condition nodes retain their original state markers. The combined operating condition state set after recovery rollback is composed of all operating condition nodes that have completed the state rollback update and the operating condition nodes that have not undergone state change. The integrated working condition state set after recovery and rollback is input into the graph state synchronization unit to construct a state conflict alignment graph. Dependency path analysis and conflict detection are performed on working condition nodes where there are state inconsistencies between the original state and the harmonized state in the multi-scale evolutionary knowledge graph. State synchronization is prioritized based on the closed-loop paths in the state conflict alignment graph, updating the state attributes of the corresponding working condition nodes in the multi-scale evolutionary knowledge graph, thus forming a harmonized integrated working condition state set, where: The construction of the state conflict alignment graph is specifically as follows: The original state label of each working condition node is read from the multi-scale evolutionary knowledge graph and compared with the state label obtained after recovery and backtracking. When the state label of the same working condition node is inconsistent in the two state sets, the working condition node is marked as a conflict node. Based on the existing time scale evolution relationship, mechanism causal relationship and cross-scale relationship in the multi-scale evolutionary knowledge graph, directed association edges are established between conflict nodes to form a state conflict alignment graph. The dependency path analysis and conflict detection are performed as follows: In the state conflict alignment diagram, the association path between conflict nodes is traversed and analyzed along the temporal evolution relationship, mechanistic causal relationship and cross-scale correlation relationship between the working condition nodes. If multiple conflict nodes form a reachable path through temporal continuity relationship and mechanistic causal chain, it is determined that the conflict states on the reachable path have a dependency relationship; otherwise, it is determined to be a local isolated conflict. The step of prioritizing state synchronization based on the closed-loop path in the state conflict alignment graph is as follows: In the state conflict alignment graph, a closed association path consisting of multiple working condition nodes is identified. For working condition nodes located on the closed loop path, the state markers in the harmonized comprehensive working condition state set are used as the benchmark to uniformly update the state attributes of the corresponding working condition nodes in the multi-scale evolutionary knowledge graph. For conflict nodes not located on the closed loop path, after the closed loop path synchronization is completed, state synchronization is performed sequentially according to the topological order in the state conflict alignment graph to avoid repeated propagation of state conflicts in the evolutionary chain, thus forming a harmonized comprehensive working condition state set consistent with the multi-scale evolutionary knowledge graph.

[0025] In this embodiment, determining the type of complex working condition and outputting the risk level and early warning information includes: Based on the harmonized comprehensive working condition state set, working condition nodes in latent, fluctuating, and dangerous states are screened to determine the target working condition node set for participating in complex working condition identification. Based on the target working condition node set, the connectivity relationships between the target working condition nodes are identified in the multi-scale evolutionary knowledge graph according to the temporal evolutionary relationships, mechanistic causal relationships, and cross-scale correlations among the working condition nodes. The set of working condition nodes with continuous correlations is identified as a complex working condition cluster, where: Continuous association refers to the fact that in a multi-scale evolutionary knowledge graph, the target working condition node can form at least one uninterrupted association path in chronological order, based on the evolutionary relationship along the time scale, the causal relationship of the mechanism, and the cross-scale association. For each complex operating condition cluster, the state distribution characteristics of each operating condition node within the cluster are analyzed. The distribution of each node across susceptible, latent, fluctuating, dangerous, and recovering states is recorded. Furthermore, the evolutionary path information corresponding to the complex operating condition cluster is obtained by combining a multi-scale evolutionary knowledge graph. Specifically, the acquisition of the evolutionary path information corresponding to the complex operating condition cluster involves: In the multi-scale evolutionary knowledge graph, the connection order of the working condition nodes in the complex working condition cluster is tracked along the established time-scale evolutionary relationships, mechanistic causal relationships and cross-scale correlations between the working condition nodes in the complex working condition cluster. According to the order of appearance of the working condition nodes at the millisecond, second, minute and hour time scales, the change trajectory of the working condition node state marker is read to form the evolution path information describing the evolution process of the working condition node state from susceptible state, latent state, fluctuating state to dangerous state and recovery state. The state distribution characteristics and evolution path information corresponding to complex operating condition clusters are correlated with standardized multi-source operational data to obtain correlation characteristics. Based on the correlation characteristics, the complex operating condition clusters are classified as complex operating condition types, and the corresponding risk levels are determined. Furthermore, early warning information is generated based on the evolution path information of the complex operating condition clusters. The acquisition of the associated features is specifically as follows: Read the state distribution characteristics of working condition nodes within a complex working condition cluster and their evolution path information in a multi-scale evolutionary knowledge graph. Extract data segments consistent with the time scale of the working condition nodes from standardized multi-source operational data. Align the state change sequence and state duration of the working condition nodes with the operational data within the corresponding time period to form a correlation feature that represents the correspondence between the evolution of working condition state and the changes in actual operational data. The determination of complex operating condition type based on correlation features for complex operating condition clusters specifically includes: Based on the distribution ratio of working conditions, the concentration of dangerous and fluctuating states, and the changing trends in evolution path information reflected in the correlation characteristics, complex working condition clusters are classified. When the proportion of dangerous states is high and the evolution path shows a continuous increasing trend, it is judged as a high-risk complex working condition. When latent and fluctuating states are the main ones and the evolution path shows intermittent changes, it is judged as a medium-risk complex working condition. When susceptible and recovery states are the main ones and the evolution path shows a stable and regressive trend, it is judged as a low-risk complex working condition.

[0026] Example 1: To verify the feasibility of this invention in practice, it was applied to a SiC power module used in a main drive inverter of a new energy vehicle. This inverter operates under conditions of alternating high-frequency switching, high load fluctuations, and complex cooling conditions. In actual operation, there were numerous instances of abnormal junction temperature fluctuations and frequent voltage overshoots without triggering protection thresholds. Traditional threshold-based monitoring methods only detect risks after significant performance degradation or protection activation, failing to identify latent complex operating conditions in advance and posing a high risk of failure.

[0027] This invention processes multi-source operating data collected during inverter operation, including electrical signal data such as switching voltage, current, and dv / dt; thermal signal data such as junction temperature, case temperature, and temperature rise rate; structural and environmental parameters such as ambient temperature and radiator status; and control behavior data such as drive frequency and load adjustment commands. Cross-scale operating condition features are constructed at millisecond, second, minute, and hourly time scales, forming a multi-scale evolutionary knowledge graph that expresses the operating condition nodes and their evolutionary relationships at each time scale.

[0028] During continuous operation testing, data collection and analysis were conducted on the inverter over a period of 30 days, accumulating approximately 720 hours of operation. From day 12 to day 18, no parameter exceedances were observed. However, the method of this invention identified multiple operating nodes in a latent state evolving towards a fluctuating state on day 13. These nodes exhibited a significant increase in the number of dv / dt peaks at the millisecond level, a gradually accelerating junction temperature rise rate at the minute level, and a slow drift in the thermal resistance estimate at the hour level. After improving the SEIR model and reconciling the spectral feedback, this stage was identified as a medium-to-high risk complex operating condition cluster, and early warning information was issued in advance.

[0029] Based on the early warning results, the maintenance personnel adjusted the cooling strategy and load scheduling, which reduced the peak junction temperature of the device by about 8.6°C and reduced the number of times the peak dv / dt exceeded the reference value by about 47%, verifying the effectiveness and practical value of the invention in actual engineering scenarios.

[0030] Table 1. Data on Complex Operating Conditions and Risk Evolution During SiC Inverter Operation

[0031] As shown in Table 1, during the continuous operation of the SiC inverter, the device operating conditions exhibit distinct phased evolution characteristics, and the changing trends of multi-source operating indicators in different operating stages show strong consistency and cumulativeity. In the initial stage of operation, the maximum junction temperature peak remained between 87.2℃ and 91.5℃, the number of times the dv / dt peak value exceeded the limit was low, and the estimated thermal resistance change rate was less than 1% / 100h, indicating a stable overall operating condition. The method of this invention identified risks in this stage, classifying it as low risk. After day 11, the number of times the dv / dt peak value exceeded the limit and the thermal resistance change rate began to rise synchronously. The method of this invention identified a latent risk state, demonstrating its ability to capture potential abnormal trends reflected by the coordinated changes of multiple indicators before obvious faults occur.

[0032] On days 13 and 15, the peak junction temperature continued to rise to 98.4℃ and 101.6℃, respectively. The number of times the peak dv / dt exceeded the limit increased significantly, and the thermal resistance change rate rose to 1.8% / 100h and 2.6% / 100h, respectively. The method of this invention upgraded the risk assessment of the operating condition at this stage from medium risk to medium-high risk, demonstrating the ability to continuously track the evolution of the operating condition. When the operation reached day 18, all indicators reached a high level, and the risk assessment level was high risk, indicating that a complex operating condition had formed and had the potential for further deterioration.

[0033] After implementing operational adjustments, data from 3 and 7 days post-adjustment show a significant decrease in peak junction temperature, number of dv / dt exceedances, and thermal resistance change rate. This indicates that the method of this invention can identify the formation process of complex operating conditions and reflect the recovery trend of the conditions after adjustment. Comprehensive analysis shows that the method of this invention can achieve early identification, dynamic evaluation, and evolution tracking of complex operating conditions of SiC devices based on multi-source operational data and time-scale evolution characteristics, verifying its effectiveness and reliability in practical engineering applications.

[0034] 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 SiC device complex working condition identification method based on a knowledge graph, characterized by, The application relates to a SiC device operation condition prediction method and device. The method comprises the following steps: collecting multi-source operation data of SiC device operation processes, and pre-processing the multi-source operation data to obtain standardized multi-source operation data; based on the standardized multi-source operation data, time scale division is carried out, cross-scale working condition features of millisecond, second, minute and hour levels are respectively constructed, and a cross-scale working condition feature set is formed by combination; based on the cross-scale working condition feature set, a multi-scale evolution knowledge graph is constructed, and the cross-scale working condition features are respectively bound to corresponding working condition nodes; based on the multi-scale evolution knowledge graph, an improved SEIR model is constructed, state mapping is carried out, the working condition nodes are marked as susceptible state, latent state, fluctuation state, dangerous state and recovery state, and a comprehensive working condition state set is formed; a graph feedback reconciliation module is constructed, semantic consistency verification, latent state amplification and recovery state rollback processing are carried out on the comprehensive working condition state set, and a reconciled comprehensive working condition state set is obtained; 2. The SiC device complex working condition recognition method based on a knowledge graph according to claim 1, characterized in that, the reconciled comprehensive working condition state set is identified in the multi-scale evolution knowledge graph, a complex working condition cluster is formed, and the complex working condition type is determined according to the state distribution, evolution path and mechanism correlation of the complex working condition cluster, and the risk level and early warning information are output.

3. The SiC device complex working condition recognition method based on a knowledge graph according to claim 1, characterized in that, The multi-source operation data specifically includes electrical signal data, thermal signal data, structure and environment parameters and control behavior data.

4. The SiC device complex working condition recognition method based on a knowledge graph according to claim 1, characterized in that, The pre-processing of the multi-source operation data specifically includes time synchronization, sampling alignment, abnormal data elimination, dimension unification, normalization conversion and denoising processing. The combination to form the cross-scale working condition feature set comprises: the standardized multi-source operation data is respectively subjected to time scale division according to millisecond, second, minute and hour levels, a data analysis window and a sliding interval between adjacent windows are set on each time scale, and each data segment after time scale division is subjected to segmentation processing; for each data segment under each time scale, cross-scale working condition features are extracted from the electrical signal data, the thermal signal data, the structure and environment parameters and the control behavior data; 5. The SiC device complex working condition recognition method based on a knowledge graph according to claim 1, characterized in that, the cross-scale working condition features extracted under each time scale are respectively corresponded to the feature set under the corresponding time scale, and the cross-scale working condition feature set is formed by combination in sequence according to millisecond, second, minute and hour levels. The construction of the multi-scale evolution knowledge graph comprises: based on the cross-scale working condition feature set, the millisecond, second, minute and hour cross-scale working condition features are respectively corresponded to millisecond working condition nodes, second working condition nodes, minute working condition nodes and hour working condition nodes, and the time scale identifier and the sampling time period identifier are recorded for each working condition node; based on the working condition nodes, a node set and a relationship set of the multi-scale evolution knowledge graph are constructed, the node set comprises working condition nodes, parameter nodes, mechanism nodes and equipment state nodes, and the relationship set comprises time sequence relationship under the same time scale, association relationship between working condition nodes and parameters, causal relationship between working condition nodes and mechanisms, spatial association relationship between working condition nodes and cross-scale association relationship between different time scales; Based on four types of time scales, millisecond-level sub-graphs, second-level sub-graphs, minute-level sub-graphs and hour-level sub-graphs are constructed, and the time sequence relationship and causal association relationship between the working condition nodes in each sub-graph are established; Cross-scale association relationships are established between different time scale sub-graphs, including cross-scale transmission relationships from millisecond level to second level, from second level to minute level, and from minute level to hour level, and cross-scale influence relationships for data segments that have cross-scale influence, forming a multi-scale evolution knowledge graph; The multi-scale evolution knowledge graph is versioned and managed, and the node set and relationship set are incrementally updated according to the newly added cross-scale working condition characteristics.

6. The SiC device complex working condition recognition method based on a knowledge graph according to claim 1, characterized in that, The corresponding working condition state set is formed, including: An improved SEIR model is constructed, which consists of a state management module, a dynamic conversion rate generation module, a feedback control module, and a state evolution calculation module; Based on the multi-scale evolution knowledge graph, the state management module is initialized, the working condition nodes in the multi-scale evolution knowledge graph are written into the state management module, and the corresponding relationship between the working condition nodes and the susceptible state, latent state, fluctuation state, dangerous state and recovery state is established, forming an initial state set; The cross-scale working condition characteristics in the cross-scale working condition characteristic set are input into the dynamic conversion rate generation module, and the state conversion rate function graph is constructed according to the cross-scale characteristics and mechanism association relationship of each working condition node, recording the state conversion path from susceptible state to latent state, from latent state to fluctuation state, from fluctuation state to dangerous state, and from dangerous state to recovery state in function form, and generating the corresponding dynamic function state conversion rate; The dynamic function state conversion rate is input into the state evolution calculation module, and the time scale scheduling stack is introduced to schedule the state update tasks of millisecond level, second level, minute level and hour level, and the initial state set is updated in different time scales to obtain the updated forward working condition state set; The forward working condition state set is input into the feedback control module, and the conditional gate rollback feedback control path from dangerous state to fluctuation state and from fluctuation state to latent state is constructed according to the control behavior and cooling reinforcement characteristics, and the state rollback update is performed to obtain the rollback working condition state set; The rollback working condition state set is written back into the state management module for state coverage and replacement storage, forming a comprehensive working condition state set, and the comprehensive working condition state set is returned to the dynamic function state conversion rate generation module for dynamic update of the dynamic function state conversion rate.

7. The SiC device complex working condition recognition method based on a knowledge graph according to claim 1, characterized in that, The comprehensive working condition state set is obtained after reconciliation, including: A graph feedback reconciliation module is constructed, which consists of a state semantic consistency verification unit, a latent risk amplification unit, a recovery rollback unit and a graph state synchronization unit; The comprehensive working condition state set and the multi-scale evolution knowledge graph are input into a state semantic consistency verification unit, a semantic support intensity graph is constructed, based on the causal chain and scale coupling relationship between the working condition nodes and the mechanism nodes, parameter nodes and historical evolution trajectories in the knowledge graph, the semantic support intensity index of each dangerous state node is calculated, and the nodes below the index threshold are screened to form a set of working condition nodes to be corrected; In the state semantic consistency verification unit, the state distribution and time evolution relationship of adjacent working condition nodes are analyzed in the multi-scale evolution knowledge graph based on the semantic support intensity graph and the working condition nodes in the set of working condition nodes to be corrected, the state of the working condition nodes in the set of working condition nodes to be corrected is adjusted, and a semantic corrected comprehensive working condition state set is generated; The semantic corrected comprehensive working condition state set is input into a latent risk amplification unit, the working condition nodes in the latent state are taken as the center, the time scale evolution relationship and mechanism causal relationship in the multi-scale evolution knowledge graph are combined, the adjacent working condition nodes meeting the preset conditions are detected, and a risk level accumulation mapping matrix is introduced, different risk factor combinations are mapped into an upgrade path from the latent state to the dangerous state, and the state label of the working condition nodes meeting the latent risk amplification conditions is updated to form a comprehensive working condition state set processed by the latent risk amplification; The comprehensive working condition state set processed by the latent risk amplification is input into a recovery rollback unit, the working condition nodes in the recovery state are selected, the time scale information in the multi-scale evolution knowledge graph and the standardized multi-source operation data are combined, a stability rollback gate stack is introduced, the stability of the working condition nodes in the recovery state is periodically judged, the node state of the unstable working condition nodes is rolled back from the recovery state to the latent state, and a comprehensive working condition state set processed by the recovery rollback is obtained; The comprehensive working condition state set processed by the recovery rollback is input into a graph state synchronization unit, the working condition nodes with inconsistent states between the original state and the harmonious state in the multi-scale evolution knowledge graph are analyzed in the dependent path and the conflict is detected, the state synchronization is preferentially performed according to the closed loop path in the state conflict alignment graph, the state attribute of the corresponding working condition node in the multi-scale evolution knowledge graph is updated, and a harmonized comprehensive working condition state set is formed.

8. The SiC device complex working condition recognition method based on a knowledge graph according to claim 1, characterized in that, The determination of the complex working condition type and the output of the risk level and the early warning information include: Based on the harmonized comprehensive working condition state set, the working condition nodes in the latent state, the fluctuation state and the dangerous state are screened to determine a target working condition node set participating in complex working condition identification; Based on the target working condition node set, the time scale evolution relationship, the mechanism causal relationship and the cross-scale association relationship between the working condition nodes in the multi-scale evolution knowledge graph are identified, the working condition node set with continuous association relationship is identified as a complex working condition cluster; For each complex condition cluster, the state distribution characteristics of each condition node in the complex condition cluster are analyzed, the distribution of each condition node between the susceptible state, the latent state, the fluctuation state, the dangerous state and the recovery state is recorded, and the evolution path information corresponding to the complex condition cluster is obtained in combination with the multi-scale evolution knowledge graph; The state distribution characteristics and evolution path information corresponding to the complex condition cluster are associated with the standardized multi-source operation data to obtain associated features, the complex condition cluster is determined according to the associated features, the corresponding risk level is determined, and the early warning information is generated based on the evolution path information of the complex condition cluster.

Images