A Method for Constructing a Fault Log Knowledge Graph for Electrical Complete Equipment

Abstract

This invention relates to the field of semantic analysis technology, specifically to a method for constructing a fault log knowledge graph for electrical complete sets of equipment. Driven by ontology attribute patterns, it extracts subthreshold degradation features of resonant circuits from the induction heating operating parameter sequence, and forms degradation feature records by defining a time window based on product model switching times. Based on the billet part number, the degradation feature records are time-series aligned with the eccentric load features of downstream forging processes, and corrected by incorporating the effects of transmission cooling, generating causal evidence records. A semantically symmetric ontology structure is constructed, defining logged fault entities and unlogged potential root cause entities. Potential root cause entity nodes are instantiated based on the degradation feature records. Directed cross-process causal relationship edges are established between potential root cause entities and process-sensitive fault entities based on the causal evidence records. The knowledge graph is closed-loop constructed through topological integrity verification and iterative update mechanisms, improving the fault tracing and knowledge representation capabilities of complex equipment.

Description

Technical Field

[0001] This invention relates to the field of semantic analysis technology, specifically to a method for constructing a fault log knowledge graph for electrical complete sets of equipment. Background Technology

[0002] Knowledge graphs, as a semantic knowledge representation and organization tool, enable the formal expression and reasoning of domain knowledge by constructing a graph structure of entity nodes and relation edges, and have been widely applied in the field of industrial equipment fault diagnosis. Existing knowledge graph construction methods for industrial equipment fault logs are typically based on ontology modeling, using fault log text as the data source. Through entity recognition and relation extraction techniques, information such as fault codes, fault phenomena, and faulty components are mapped to semantic entity nodes in the graph, and semantic relationships such as causality and subordination are constructed between entities with fault log records.

[0003] The above-mentioned methods share a common underlying assumption in their ontology design: fault log records are a prerequisite for the representation of fault entity nodes in the knowledge graph. However, due to the physical mechanism of material thermodynamic state transfer between processes, the subthreshold degradation state of upstream induction heating electrical equipment can constitute the true root cause of fault logs in downstream forging and trimming equipment. This degradation state does not trigger any electrical system alarms and does not generate any fault log records. Existing knowledge graph construction methods, constrained by the above assumptions, cannot represent such unlogged root cause entities at the ontology layer, nor can they construct cross-process causal relationship edges from logged symptom nodes to unlogged root cause nodes at the relationship layer. This results in a systematic lack of root cause entities and structural errors in causal relationships in the constructed knowledge graph, failing to effectively support the root cause localization and preventive maintenance decisions for precision forging quality anomalies.

[0004] To address this, a method for constructing a fault log knowledge graph for complete sets of electrical equipment is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a method for constructing a knowledge graph of fault logs for complete sets of electrical equipment. By extracting subthreshold degradation features in the induction heating process and combining them with secondary tracking of blank parts, cross-process causal evidence is generated. Under a semantically symmetric ontology structure, a causal relationship between potential root cause entities and fault entities is constructed, and the iterative closure construction of the knowledge graph is achieved through topological integrity verification.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for constructing a fault log knowledge graph for electrical complete sets of equipment includes: Driven by the intrinsic attribute mode and with the product model switching time as the window boundary, the subthreshold degradation characteristic record of the resonant circuit is extracted from the induction heating operation parameter sequence; the intrinsic attribute mode includes the inverter operating frequency control command sequence, the reactive power compensation circuit switching action record sequence, and the inverter output side current harmonic analysis result sequence. Based on the order of billet parts exiting the furnace, the degradation feature records are aligned with the eccentric load features of the downstream forging process at the part-level time sequence. After correction for the effects of transmission and cooling, cross-process causal evidence records are generated. Construct a semantically symmetric ontology structure, define logged fault entity classes and unlogged potential root cause entity classes as mutually exclusive subclasses with semantically symmetric attribute patterns, and instantiate unlogged potential root cause entity nodes according to degradation characteristics. Based on cross-process causal evidence records, directed cross-process causal relationship edges are constructed between unlogged potential root cause entity nodes and process-sensitive logged fault entity nodes; classification topology integrity verification is performed on process-sensitive fault entities, and an iterative update process is triggered for isolated symptom nodes to complete the knowledge graph closure construction.

[0007] Preferably, the subthreshold degradation feature record of the resonant circuit is extracted from the induction heating operating parameter sequence, specifically including: Collect the body attribute mode from the induction heating control system; define the window boundary based on the product model switching log, and divide the continuous production cycle of the same model into time windows according to shifts; Within each time window, with the window start value as the reference, the cumulative change of the frequency control command within the window, the deviation between the current window reactive power compensation switching frequency and the historical reference switching frequency of the same model, and the difference between the current window current harmonic analysis mean and the historical reference mean of the same model are calculated to form a combination of degradation characteristic components. When determining subthreshold degradation, the upper and lower limits of the determination interval are determined based on the arithmetic mean and standard deviation of the historical degradation feature components. When the combination of degradation feature components falls into the upper and lower limits of the interval and shows a monotonic trend drift, it is determined to be subthreshold degradation. The degradation feature components are combined with the start and end timestamps of the corresponding window and the product model identifier and encapsulated into a complete degradation feature record.

[0008] Preferably, the steps for generating cross-process causal evidence records include: A pre-established correlation between product models and billet specifications is established, including cross-sectional area and length-to-diameter ratio. Based on historical thermal imaging measurement data of each billet model, the reference component of eccentric load caused by furnace cooling under different transmission delay conditions is calibrated. Using the billet batch number as the primary key, the degradation feature record of the shift time window to which the batch belongs is extracted from the feature database, and the eccentric load feature value of that batch is extracted. The cooling correction reference value is retrieved using the product model and the measured transmission delay as a joint index. The eccentric load feature value is subtracted from the reference value to obtain the corrected eccentric load feature component. When the corrected eccentric load characteristic component exceeds the judgment threshold and a fault log record exists for the item, a causal evidence record is generated.

[0009] Preferably, a semantically symmetric ontology structure is constructed, specifically including: In the knowledge graph ontology file, under the top-level fault-related entity parent class, logged fault entity subclasses and unlogged potential root cause entity subclasses are created through the subclass axiom, and the two are declared to be mutually exclusive through the disjoint class axiom. Define fault timestamp attribute, fault severity level attribute, device identifier attribute, product model attribute, and log text attribute for the fault entity subclass with logs; define semantically symmetrical attributes for the potential root cause entity subclass without logs: the degradation window start and end timestamp attribute is symmetrical to the fault timestamp attribute, the degradation severity description attribute is symmetrical to the fault severity level attribute, the device identifier attribute and product model attribute are completely isomorphic to the fault entity subclass with logs, and the degradation feature vector attribute stores the combined value of feature components and is symmetrical to the log text attribute.

[0010] Preferably, the instantiation of unlogged potential root cause entity nodes based on degradation feature records specifically includes: traversing degradation feature records in the feature database and filtering records that simultaneously meet the following conditions: the degradation feature component combination meets the subthreshold degradation judgment condition, and there is at least one downstream forging process fault log record within the corresponding shift time window; for each degradation feature record that meets the conditions, an unlogged potential root cause entity instance is created, and the degradation window start and end timestamps, degradation severity description strings generated according to feature component combination values, equipment identifier, product model, and degradation feature vector values ​​are sequentially filled into the corresponding ontology attributes; Register the entity instances with completed attribute population to the knowledge graph node set and establish a mapping relationship between instance identifiers and feature database record indexes.

[0011] Preferably, the determination of process-sensitive log fault entity nodes and the construction of cross-process causal relationship edges specifically include: A lookup table of fault codes and fault categories is pre-established. Overload, forging dimensional deviation, and impact load abnormality are marked as process-sensitive, while electrical short circuit and sensor disconnection are marked as equipment body type. During the instantiation of log fault entities, the fault category attribute is filled into the log fault entity instance according to the lookup table. Based on the causal evidence records in the evidence database, a directed cross-process causal relationship edge is created between the process-sensitive logged fault entity instance associated with each record and the corresponding unlogged potential root cause entity instance.

[0012] Preferably, performing classification topology integrity verification and triggering an iterative update process for process-sensitive fault entities specifically includes: For all fault category attribute nodes with log faults in the knowledge graph, count the number of isolated nodes that do not have any incoming cross-process causal relationship edges, and calculate the proportion of isolated nodes to the total number of process-sensitive fault entity nodes. When the proportion of isolated nodes exceeds the preset integrity threshold, a root cause missing alarm report is generated, which includes a list of isolated nodes and the corresponding item information and shift time window identifier for each node, and is submitted for review. The audit conclusions are written back to the knowledge graph through the ontology extension interface, triggering the generation of supplementary evidence records for the affected items and the construction of corresponding cross-process causal relationship edges; After the supplementary construction is completed, the full classification topology integrity verification is re-executed until the proportion of isolated nodes meets the preset integrity threshold, thus completing the iterative closure of the knowledge graph.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention extracts subthreshold degradation features of the resonant circuit from the induction heating operating parameter sequence and constructs degradation feature records with the product model switching time as the window boundary. This allows the system to identify potential abnormal trends before the equipment generates explicit fault logs. This invention can utilize multi-source operating data, such as frequency control command changes, reactive power compensation switching frequency deviations, and differences in current harmonic characteristics, to construct degradation feature combinations, thereby capturing implicit signals of equipment performance degradation in advance. It achieves a structured expression of the "sub-healthy state," enabling potential fault information to enter the knowledge graph in entity form, providing a more complete data foundation for subsequent fault tracing and intelligent diagnosis, and improving the equipment's operational status perception and early warning capabilities.

[0014] 2. This invention establishes a secondary tracking mechanism for billet parts, aligning the degradation feature records extracted from the upstream induction heating stage with the eccentric load features of the downstream forging process in a temporal sequence. It also combines this with corrections for the effects of transmission cooling to generate cross-process causal evidence records, thus achieving causal relationship modeling across process flows. This invention establishes the correlation between processes using information such as part number, product model, and transmission delay, thereby identifying implicit causal chains that cause anomalies in subsequent processes due to changes in the preceding process state. This mechanism not only improves the accuracy of fault tracing but also reveals the coupling effects between different production stages, providing a reliable basis for system-level fault analysis in complex manufacturing processes.

[0015] 3. This invention constructs a semantically symmetric knowledge graph ontology structure, defining "logged fault entities" and "unlogged potential root cause entities" as mutually exclusive but semantically symmetric subclasses. This allows explicit fault information and implicit degradation information to be expressed in a unified knowledge structure. Simultaneously, a classification topology integrity verification mechanism is used to perform structural detection on process-sensitive fault nodes and triggers an iterative update process for isolated symptom nodes, achieving dynamic completion and closure construction of the knowledge graph. It can automatically calculate the proportion of isolated nodes and generate root cause missing alarms, driving evidence record supplementation and causal relationship edge updates, thereby improving the structural integrity and reasoning ability of the knowledge graph and enhancing the reliability and scalability of the fault diagnosis knowledge base. Attached Figure Description

[0016] Figure 1 A schematic diagram of a method for constructing a fault log knowledge graph for complete sets of electrical equipment provided by the present invention; Figure 2 This is a schematic diagram of the logic flow for generating cross-process causal evidence provided by the present invention; Figure 3 A schematic diagram of the topology integrity verification and iterative closure logic flow provided by the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0018] Example 1:

[0019] Please see Figure 1This invention provides a method for constructing a fault log knowledge graph for complete electrical equipment. The technical solution is as follows: Driven by an ontological attribute pattern and using the product model switching time as the window boundary, subthreshold degradation feature records of the resonant circuit are extracted from the induction heating operating parameter sequence. The ontological attribute pattern includes the inverter operating frequency control command sequence, the reactive power compensation circuit switching action record sequence, and the inverter output side current harmonic analysis result sequence. All three sequences are collected at the furnace batch as the smallest granularity and retain timestamps. Using the billet batch exit sequence as the tracking basis, the degradation feature records are compared with the eccentric load characteristics of the downstream forging process. Secondary temporal alignment is performed, and cross-process causal evidence records are generated after correction for transmission cooling effects. A semantically symmetric ontology structure is constructed, defining logged fault entity classes and unlogged potential root cause entity classes as mutually exclusive subclasses with semantically symmetric attribute patterns. Unlogged potential root cause entity nodes are instantiated based on degradation feature records. Based on cross-process causal evidence records, directed cross-process causal relationship edges are constructed between unlogged potential root cause entity nodes and process-sensitive logged fault entity nodes. Classification topology integrity verification is performed on process-sensitive fault entities, and an iterative update process is triggered for isolated symptom nodes to complete the knowledge graph closure construction.

[0020] Furthermore, the process of obtaining the degradation feature record specifically includes: The body attribute modes are collected separately from the induction heating control system; the window boundaries are defined based on the product model switching log, and the continuous production cycle of the same model is divided into several time windows according to the shift, so as to ensure that each time window contains only the production data of the same product model. Within each time window, with the window start value as the reference, the cumulative change of the frequency control command within the window, the deviation between the current window reactive power compensation switching frequency and the historical reference switching frequency of the same model, and the difference between the current window current harmonic analysis mean and the historical reference mean of the same model are calculated. These three together constitute the degradation characteristic component combination. When determining subthreshold degradation, the lower limit of the determination interval is first determined: select historical normal production data from at least 10 shifts, calculate the arithmetic mean and standard deviation of each characteristic component, and use the lower limit equal to the mean plus twice the standard deviation as the upper limit of normal fluctuation; secondly, the upper limit of the determination interval is determined: trace back the characteristic average values ​​from at least 3 shifts before the historical fault occurred, and take the minimum value among them as the upper limit reference. When a characteristic component falls between the lower and upper limits and exhibits a monotonically increasing or decreasing trend drift, it is determined to be subthreshold degradation.

[0021] Based on the upper and lower limits of the judgment interval, the degradation feature components are combined with the start and end timestamps of the corresponding window and the product model identifier to encapsulate a complete degradation feature record, which is then written into the feature database using the shift time window and the product model label as dual indexes.

[0022] Specifically, the induction heating control system continuously adjusts the inverter's operating frequency to maintain the resonant circuit near its resonant point for maximum energy transfer efficiency. When the capacitance in the resonant circuit drifts due to long-term operation, or when the electromagnetic coupling parameters between the coil and the billet shift due to furnace lining erosion, the inverter's frequency control command needs to be continuously adjusted in the compensation direction to maintain the resonant state. This cumulative adjustment is small within a single production cycle and does not trigger any alarms, but it exhibits a monotonic deviation trend within a continuous production window of the same model, carrying information about the deterioration of the resonant circuit's health. Using the frequency control command value at the beginning of the window as a reference, the cumulative change in the frequency control command within the window is calculated to eliminate the influence of differences in process parameters between different product models on the reference frequency, ensuring that the cumulative change only reflects the relative deviation within the window period. To avoid the window's initial value itself losing its reference value due to degradation at the end of the previous window, the distribution range of the initial value of each window is compared with the historical normal initial value of the same model during the statistical analysis phase. If the initial value has significantly deviated from the historical normal range, the validity of the reference benchmark for that window is marked in the degradation feature record for reference in the subsequent entity instantiation phase.

[0023] The reactive power compensation circuit compensates for changes in reactive power demand during induction heating by switching capacitor banks. When the resonant parameters degrade, the switching frequency of the reactive power compensation circuit deviates from the reference frequency during normal production of the same model. By using the historical reference switching frequency of the same model as a reference, the influence of differences in reactive power compensation demand between different product models is eliminated, ensuring that the switching frequency deviation only reflects changes in equipment status rather than differences in process parameters.

[0024] Changes in the harmonics of the inverter output current reflect alterations in the nonlinear characteristics of the resonant circuit. When capacitors age or contact resistance increases, the quality factor of the resonant circuit decreases, and the proportion of non-fundamental components in the output current increases. By using the historical average value of the same model as a reference, the differences in the influence of different billet models on the resonant characteristics are eliminated, ensuring that the harmonic mean difference only reflects the degradation state of the equipment itself.

[0025] After the equipment is put into use, from the historical operating data of continuous production of the same model, a historical period confirmed by equipment maintenance records to be in normal operating condition is selected. The upper bound of the distribution of the cumulative change in frequency control commands, the frequency deviation of reactive power compensation switching, and the mean difference of current harmonics under normal conditions is statistically analyzed and used as the lower limit of the subthreshold degradation judgment interval. From the most recent consecutive shift time windows before the recorded equipment fault, historical values ​​of three characteristic components are extracted. The lowest value among them with an observable obvious trend shift is used as the reference upper bound of the judgment interval. The subthreshold degradation judgment interval is determined between the lower limit and the reference upper bound. When the value of at least one component in the degradation characteristic component combination falls into this judgment interval, the window is determined to have subthreshold degradation. The specific interval boundary values ​​are confirmed by engineers after statistical analysis of the historical data of this equipment model, stored in the configuration table of the characteristic database, and can be updated as the accumulated data increases.

[0026] The encapsulation format of the degradation feature record includes six fields: window start timestamp, window end timestamp, product model identifier, cumulative change in frequency control command, reactive power compensation switching frequency deviation, and mean difference in current harmonics. The first three fields are used for subsequent timing alignment and cross-process association, while the latter three fields constitute the degradation feature vector, which is used for subsequent entity node attribute filling.

[0027] Furthermore, after correction for the effects of transmission cooling, a record of cross-process causal evidence is generated for reference. Figure 2 Specifically, it includes: A correlation between product model and billet specifications is established in advance. The billet specifications include cross-sectional area and length-to-diameter ratio. Based on historical thermal imaging measurement data of each billet model, the reference component of eccentric load caused by furnace cooling under different transmission delay conditions is calibrated. The calibrated value comprehensively reflects the actual superposition effect of heat dissipation on the billet surface and internal longitudinal heat conduction. It is stored in the correction reference database according to the combination of model and transmission delay. Using the billet part number as the primary key, the degradation feature record of the shift time window to which the part belongs is extracted from the feature database, and the eccentric load feature value of the part is extracted from the forging process control system. The cooling correction reference value is retrieved by using the product model and the measured transmission delay as a joint index. The corrected eccentric load characteristic component is obtained by subtracting the reference value from the eccentric load characteristic value. When the corrected eccentric load characteristic component exceeds the judgment threshold and there is a fault log record for that incident, a causal evidence record is generated, which includes the incident number, the combined value of the upstream degradation characteristic components, the corrected eccentric load characteristic component value, the measured transmission delay, and the identifier of the associated fault log record.

[0028] Specifically, the correspondence between product models and billet specifications is pre-established and stored in the configuration database. The cross-sectional area determines the heat capacity of the billet during transport. A billet with a larger cross-sectional area has a relatively smaller temperature difference between the surface and the core under the same transport time, resulting in a relatively lower reference component of the eccentric load caused by cooling. The length-to-diameter ratio determines the attitude stability of the billet on the roller conveyor and the heat dissipation ratio of each part. A billet with a larger length-to-diameter ratio generates an axial temperature gradient during transport due to the different heat dissipation area ratio between the ends and the middle, forming the axial component of the eccentric load.

[0029] Production sections with confirmed normal induction heating quality and normal forging equipment status were selected from historical production data. Using product model and measured transmission delay as a joint index, the transmission delay was divided into multiple intervals based on the roller conveyor speed adjustment range. The granularity of the division ensured that the difference in cooling effect within the same interval was within the accuracy range of the eccentric load measurement. The mean value of the eccentric load characteristic value under each joint index combination was statistically analyzed. This mean value served as the reference component of the eccentric load caused by cooling factors within the corresponding transmission delay interval for the corresponding model, and was stored in the reference component database. The eccentric load characteristic value was acquired by a pressure sensor installed on the hydraulic system at a sampling frequency of 50 Hz and processed by a second-order low-pass filter for noise reduction.

[0030] When cooling correction is applied to the eccentric load characteristic value, the reference component value corresponding to the measured transmission delay is matched. The original eccentric load characteristic value is subtracted from the reference component value to obtain the corrected eccentric load characteristic component. This component eliminates the influence of transmission cooling factors and only reflects the residual eccentric load level determined by the upstream heating quality and the forging process itself.

[0031] The cooling correction baseline value is retrieved using the product model and measured transmission delay as a joint index. If the measured transmission delay falls between two calibration intervals, cubic spline interpolation is used to obtain the correction baseline to compensate for the nonlinearity of the heat conduction process. The baseline value is subtracted from the eccentric load characteristic value to obtain the corrected eccentric load characteristic component. When the corrected eccentric load characteristic component exceeds the judgment threshold and a fault log record exists for that incident, a causal evidence record is generated.

[0032] The eccentric load judgment threshold is obtained as follows: During a historical period when the induction heating quality is normal and the forging equipment is in normal condition, the distribution range of the corrected eccentric load characteristic components is statistically analyzed, and the upper limit of this distribution range is used as the benchmark reference value. Combining the measurement uncertainty of the eccentric load with the allowable process fluctuation range of the forging process, the benchmark reference value is appropriately corrected upwards to obtain the initial setting value of the judgment threshold. Then, the measured values ​​of the corrected eccentric load characteristic components in historical eccentric load anomaly events confirmed by manual tracing as being caused by upstream heating degradation are used to verify the ability of the initial setting value to distinguish historically known anomalies. If necessary, the initial setting value is adjusted. Judgment thresholds are set separately using product models as an index, because different models of billets have different load sensitivities during the forging process. Specific thresholds are confirmed by engineers based on the above statistical results and stored in the configuration database, which can be reviewed and updated as data accumulates.

[0033] When the corrected eccentric load characteristic component exceeds the judgment threshold for the corresponding model, and there is at least one fault record in the fault log of the forging process for that part, a causal evidence record is generated. The causal evidence record includes the following fields: part number, corresponding shift time window identifier, degradation characteristic record index, fault log record index, and corrected eccentric load characteristic component value.

[0034] Furthermore, a semantically symmetric ontology structure is constructed, specifically including: In the knowledge graph ontology file, under the top-level fault-related entity parent class, logged fault entity subclasses and unlogged potential root cause entity subclasses are created through the subclass axiom, and the two are declared to be mutually exclusive through the disjoint class axiom. Define fault timestamp attribute, fault severity level attribute, device identifier attribute, product model attribute, and log text attribute for the log fault entity subclass; Define semantically symmetrical attributes for the non-logged potential root cause entity subclass: the degradation window start and end timestamp attribute is symmetrical to the fault timestamp attribute; the degradation severity description attribute is symmetrical to the fault severity level attribute; the device identifier attribute and product model attribute are completely isomorphic to the logged fault entity subclass; the degradation feature vector attribute stores the feature component combination value and is symmetrical to the log text attribute. The semantically symmetric structure enables the inference engine to traverse two types of entity nodes simultaneously using a unified query rule, without needing to distinguish the subclass of the entity in the rule, thus providing an ontology-level guarantee for the semantic consistency of cross-type causal relationship edges.

[0035] Specifically, the knowledge graph ontology file is constructed using a standard ontology description language. Under the top-level fault-related entity parent class, logged fault entity subclasses and non-logged potential root cause entity subclasses are created respectively through subclass axioms; the two are declared to be mutually exclusive through disjoint class axioms, ensuring that no specific entity instance can be classified into both subclasses at the same time, thereby maintaining the consistency of classification during graph reasoning and querying.

[0036] The attributes of the log fault entity subclass are defined as follows: Fault Timestamp attribute, data type is timestamp, records the time when the fault event was recorded by the system; Fault Severity Level attribute, data type is enumerated string, the value is a predefined severity level vocabulary, reflecting the degree of impact of the fault on equipment operation; Equipment Identifier attribute, data type is string, records the specific equipment number where the fault occurred; Product Model attribute, data type is string, records the product model that was being produced at the time the fault occurred; Log Text attribute, data type is string, the fault description text automatically generated by the storage device management system or filled in by maintenance personnel.

[0037] The attributes of the unlogged potential root cause entity subclass are semantically symmetrical with those of the logged fault entity subclass, specifically as follows: the degradation window start timestamp attribute and the degradation window end timestamp attribute are jointly symmetrical with the fault timestamp attribute, both used to locate events on the timeline. Logged entities are located at a single point in time, while unlogged entities are located at a time interval. The two are semantically equivalent in terms of time location. The degradation severity description attribute is symmetrical with the fault severity level attribute, both describing the degree to which the equipment status deviates from the normal level. The equipment identification attribute and product model attribute are completely isomorphic with their corresponding attributes in the logged fault entity subclass, with the same value range and format definition. The degradation feature vector attribute is symmetrical with the log text attribute. Both are descriptive records of the abnormal phenomena of the equipment represented by the entity. Logged entities describe them in free text form, while unlogged entities store the values ​​of three feature components—cumulative change in frequency control commands, reactive power compensation switching frequency deviation, and mean difference of current harmonics—in the form of a fixed-length numerical sequence.

[0038] Semantic symmetry refers to the logical mapping relationship between the attributes of two types of entities in terms of temporal location, severity, and object identification. In causal reasoning, the decision logic requires that the fault time must fall within a preset maximum delay time after the opening and closing of the degradation window. This delay time is usually set to 30 to 120 seconds to reflect the billet transfer process.

[0039] The purpose of semantic symmetric design is to enable upper-layer applications of the graph (such as fault tracing reasoning and similar fault retrieval) to use a unified attribute access interface when traversing two types of nodes, thereby reducing the development complexity of upper-layer applications and ensuring the consistency of cross-type node comparison analysis.

[0040] Furthermore, based on the degradation characteristics, instantiate potential root cause entity nodes without logs, specifically including: Traverse the degraded feature records in the feature database and filter the records that meet the following conditions: the combination of degraded feature components meets the subthreshold degradation judgment condition, and there is at least one downstream forging process fault log record in the corresponding shift time window. For each degradation feature record that meets the conditions, create a logless potential root cause entity instance and fill the corresponding ontology attributes with the degradation window start and end timestamps, the degradation severity description string generated by the feature component combination value, the device identifier, the product model and the degradation feature vector value in sequence. Register the entity instances with completed attribute population to the knowledge graph node set, and establish a mapping relationship between instance identifiers and feature database record indexes to support source tracing queries when constructing causal relationship edges.

[0041] Specifically, all degraded feature records stored in the feature database are traversed, and a double screening is performed on each record: First, the combination of degraded feature components satisfies the subthreshold degradation judgment condition, that is, the value of at least one feature component falls within the subthreshold degradation judgment interval; second, within the shift time window corresponding to the degraded feature record, there exists at least one fault log record in the fault log database of the downstream forging process. Both conditions must be met simultaneously for the degraded feature record to be used as the basis for instantiating a potential root cause entity node without logs.

[0042] Unlogged potential root cause nodes are instantiated only when there are fault log records downstream. The purpose is to control the node size of the knowledge graph within a range that is relevant to faults: if a purely subthreshold degradation state does not produce any observable fault consequences downstream within the corresponding time window, it will not be included in the graph for the current construction purpose (fault tracing and causal link construction) to avoid generating a large number of isolated degradation nodes that cannot be associated with any faults, which would affect the query efficiency and knowledge density of the graph.

[0043] For each degradation feature record that meets the dual screening criteria, create a logless potential root cause entity instance. The attributes are filled according to the following rules: the degradation window start and end timestamps are directly taken from the window start and end timestamp fields of the degradation feature record; the degradation severity description string is generated based on the feature component combination values ​​according to the following rules: the value of each feature component is compared with the width of its corresponding subthreshold degradation judgment interval. The interval is divided into three equal segments based on the width of the judgment interval. Those falling into the lower segment correspond to a "slight degradation" description, those falling into the middle segment correspond to a "moderate degradation" description, and those falling into the higher segment or having two or more feature components simultaneously entering the judgment interval correspond to a "significant degradation" description. The above segment boundaries can be adjusted by engineers in the configuration file according to the model. The equipment identifier and product model are directly taken from the corresponding fields of the degradation feature record; the degradation feature vector attributes store three component values ​​in the order of cumulative change in frequency control instructions, reactive power compensation switching frequency deviation, and current harmonic mean difference.

[0044] After the attribute population is completed, the entity instance is registered to the knowledge graph node set. At the same time, the mapping table records the correspondence between the instance identifier and the primary key of the degenerate feature record in the feature database, so as to quickly locate the corresponding unlogged node when processing causal evidence records in the future.

[0045] Furthermore, the determination of process-sensitive log fault entity nodes and the construction of cross-process causal relationship edges specifically include: A lookup table of fault codes and fault categories is pre-established. Overload, forging dimensional deviation, and impact load abnormality are marked as process-sensitive, while electrical short circuit and sensor disconnection are marked as equipment body type. During the instantiation of log fault entities, the fault category attribute is filled into each log fault entity instance according to the lookup table. Based on the causal evidence records in the evidence database, a directed cross-process causal relationship edge is created between the process-sensitive logged fault entity instance associated with each record and the corresponding unlogged potential root cause entity instance. The attributes of the edge include a material transfer path description string, the inter-process part delay value, and the causal evidence record identifier. No cross-process causal relationship edges are constructed for equipment-type fault entities.

[0046] Specifically, a table mapping fault codes to fault categories is pre-established and stored in the configuration database. The table is established based on the following: Overload faults are triggered by mechanical loads exceeding the rated load capacity of the equipment. Local overload caused by eccentric forging is a significant source of this type of fault, and its severity is directly related to the uniformity of upstream billet temperature distribution, thus it is classified as process-sensitive. Forging dimensional deviation faults reflect uneven billet filling during forging. Differences in material fluidity caused by uneven temperature distribution are an important influencing factor, thus it is classified as process-sensitive. Abnormal impact load faults are related to the impact force distribution when the billet enters the mold. Local hardness differences caused by uneven billet temperature will exacerbate the unevenness of impact load, thus it is classified as process-sensitive. Electrical short circuit faults are caused by electrical system insulation failure or circuit abnormalities, and are not directly related to the upstream billet heating quality, thus it is classified as equipment-related. Sensor disconnection faults are caused by abnormal sensor physical connections, and are also unrelated to the upstream process status, thus it is classified as equipment-related.

[0047] During the instantiation phase of entities with log faults, for each entity instance with log faults, the fault code lookup table is consulted, and the fault category attribute value obtained is filled into the fault category attribute field of the entity instance.

[0048] The process of constructing cross-process causal relationship edges is as follows: Traverse all causal evidence records in the evidence database. For each record, locate the corresponding logged fault entity instance through the fault log record index in the record, and determine whether its fault category attribute is process-sensitive. If it is process-sensitive, locate the corresponding unlogged potential root cause entity instance through the degradation feature record index in the record, and create a directed cross-process causal relationship edge between the two. The direction of the edge is from the unlogged potential root cause entity node to the logged fault entity node. Attach the primary key of the causal evidence record to the edge as a tracing identifier so that the original evidence supporting the causal relationship can be traced in subsequent queries. If the fault category attribute is equipment body type, skip the causal evidence record and do not construct a causal relationship edge.

[0049] Furthermore, for process-sensitive fault entities, a classification topology integrity verification is performed and an iterative update process is triggered, specifically including: For all fault category attribute nodes with log faults in the knowledge graph, count the number of isolated nodes that do not have any incoming cross-process causal relationship edges, and calculate their proportion to the total number of process-sensitive fault entity nodes. When the proportion of isolated nodes exceeds the preset integrity threshold, a root cause missing alarm report is generated, which includes a list of isolated nodes and the corresponding part information and shift time window identifier for each node, and is submitted to the process engineer for review. The process engineer's review conclusion is written back to the knowledge graph through the ontology extension interface, triggering the generation of supplementary evidence records for the affected parts and the construction of corresponding cross-process causal relationship edges; After the supplementary construction is completed, the full classification topology integrity verification is re-executed until the proportion of isolated nodes meets the preset integrity threshold, thus completing the iterative closure of the knowledge graph. Figure 3 .

[0050] Specifically, integrity verification is performed after all logged faulty entities and unlogged potential root cause entities have been instantiated and all cross-process causal relationship edges have been initially constructed.

[0051] For all logged fault entity nodes in the knowledge graph with the fault category attribute value of "process sensitive," traverse and count the nodes that do not have any incoming cross-process causal relationship edges. That is, process-sensitive fault entity nodes that are not pointed to by any unlogged potential root cause entity node through causal relationship edges are called isolated symptom nodes. Calculate the proportion of isolated symptom nodes to the total number of process-sensitive fault entity nodes, using this as the isolated node proportion indicator for the current integrity verification. Re-execute the verification after supplementary construction. If the isolated node proportion decreases by less than 2% in two consecutive iterations, or the total number of iterations reaches the upper limit of 10, stop the iteration and complete the knowledge graph closure.

[0052] The preset integrity threshold is obtained as follows: Before the knowledge graph is put into use, based on the results of manual source tracing analysis of historical fault data, the historical proportion of process-sensitive faults that have been manually confirmed to have upstream degradation causes is statistically analyzed. Combining the current sensor layout density on the production line and the integrity of historical data collection, the above historical proportion is appropriately reduced to obtain an initial reference value for the preset integrity threshold. The reduction range is estimated and confirmed by engineers based on the on-site sensor coverage and the historical data loss rate. The specific threshold is stored in the configuration file and can be reviewed and updated as the knowledge graph accumulates data.

[0053] When the proportion of isolated nodes exceeds the preset integrity threshold, the system generates a root cause missing alarm report, which includes a list of isolated symptom nodes (each node includes an entity identifier, fault code, and fault timestamp), the part number, billet specifications, measured transmission delay, and shift time window identifier for each isolated node, and submits it to the manual review stage.

[0054] The audit conclusion is written back to the knowledge graph through the ontology extension interface, triggering the following supplementary construction process: If the audit confirms that the fault corresponding to the isolated node has an upstream degradation cause that has not yet been covered by sensor data, a corresponding unlogged potential root cause entity instance is generated. If the corresponding degradation feature record is missing due to data collection issues, the engineer will supplement the feature data based on manual records before instantiation. Subsequently, cross-process causal relationship edges are constructed between the newly created unlogged potential root cause entity instance and the corresponding isolated symptom node, and the corresponding causal evidence record in the evidence database is updated.

[0055] After the supplementary construction is completed, the full classification topology integrity verification is re-executed, and the new isolated node ratio index is calculated. If it still exceeds the preset integrity threshold, the iterative process of generating alarm reports, manual review, and supplementary construction is repeated until the isolated node ratio drops below the preset integrity threshold, and the iterative closure construction of the knowledge graph is completed.

[0056] Example 2:

[0057] This embodiment, based on Embodiment 1, introduces a weighted composite degradation scoring mechanism based on the correlation of historical data in the subthreshold degradation determination stage, and performs cross-window degradation trend extraction and degradation trajectory temporal edge construction after the knowledge graph node instantiation is completed. The process of collecting induction heating operation parameters, window division, degradation feature component extraction, cross-process causal evidence record generation, semantic symmetric ontology structure construction, and classification topology integrity verification is the same as in Embodiment 1.

[0058] In the subthreshold degradation assessment stage, a weighted composite degradation score based on the correlation of historical data is used instead of independent single-component assessment. The specific steps include: Historical degradation feature records of the same equipment and product model are extracted from the feature database in advance. The offsets of the three feature components within each shift time window are used as inputs, and whether fault log records are generated in the downstream forging process during the corresponding shift time window are used as labels. The absolute values ​​of the absolute values ​​of the offsets of the three feature components—cumulative change of frequency control command, frequency deviation of reactive power compensation switching, and mean difference of current harmonics—are calculated and compared with the absolute values ​​of the fault labeling results. The absolute values ​​of the correlation of the three components are normalized to obtain their respective weighting coefficients, and the sum of the weighting coefficients is one. When the absolute value of the correlation of a certain component is lower than the minimum effective correlation specified in the configuration file, the weighting coefficient of that component is set to zero and the remaining components are renormalized. In the degradation judgment stage of each time window, the difference between each of the three feature components and the lower limit of the corresponding subthreshold degradation judgment interval is divided by the width of the corresponding judgment interval to obtain the dimensionless offset. Each dimensionless offset is multiplied by the corresponding weighting coefficient and then summed to obtain the composite degradation score of the current window. When the composite degradation score exceeds the composite score judgment threshold determined by the same statistical method, it is determined that there is subthreshold degradation in the window. The weighting coefficients are stored in the configuration database with the product model as the index, and are updated regularly as the data in the feature database accumulates.

[0059] Specifically, the subthreshold degradation determination uses a logical OR judgment based on at least one of the three feature components falling within the determination interval, without distinguishing the actual diagnostic sensitivity differences of each component to the degradation state of different billet types. This embodiment replaces the above judgment method with a weighted composite degradation score. In the subthreshold degradation determination stage, a weighted composite degradation score based on the correlation of historical data is used. First, Pearson correlation analysis is used to calculate the correlation coefficient between the offset of each feature component and the downstream fault label. If the absolute value of the correlation coefficient of a certain component is lower than the 90th percentile value based on the normal distribution assumption, it is determined that it has no diagnostic significance, and its weight is set to zero.

[0060] In the degradation determination stage of each time window, the lower limit of each feature component is subtracted and then divided by the determination interval width to obtain a dimensionless offset between 0 and 1. Each dimensionless offset is multiplied by its corresponding weighting coefficient and then summed to obtain the composite degradation score of the current window. When the composite degradation score exceeds a preset threshold, it is determined that the window has subthreshold degradation.

[0061] The weighting coefficients are obtained as follows: Historical degradation feature records of the same equipment and product model are extracted from the feature database. The absolute values ​​of the offsets of three feature components—cumulative change in frequency control commands, reactive power compensation switching frequency deviation, and mean difference in current harmonics—are used as inputs. Whether a fault log record was generated in the downstream forging process during the corresponding window is used as a label. The absolute values ​​of the correlation between the absolute values ​​of the offsets of the three components and the fault labeling results are calculated separately. The minimum effective correlation is initially referenced by the upper bound of the distribution of the absolute values ​​of the correlation between the offsets of each component and the fault labeling in the historical normal production section. This is confirmed by engineers and stored in the configuration file. When the absolute value of the correlation of a component is lower than this minimum effective correlation, the weighting coefficient of that component is set to zero, and the absolute values ​​of the correlation of the remaining components are renormalized so that the sum of the weighting coefficients is one. The weighting coefficients are stored in the configuration database indexed by the product model and are updated periodically as data accumulates.

[0062] The composite degradation score is calculated as follows: the difference between each of the three feature components and the lower limit of the corresponding subthreshold degradation judgment interval is divided by the width of the corresponding judgment interval to obtain the dimensionless offset of each component, thus eliminating the influence of the difference in dimensions of different components; each dimensionless offset is multiplied by the corresponding weighting coefficient and then summed to obtain the composite degradation score of the current window.

[0063] The composite degradation score threshold is obtained as follows: The upper bound of the composite degradation score distribution during historical normal operation is used as the baseline reference value. Combined with the measured composite degradation scores corresponding to windows historically confirmed by manual traceability to exhibit sub-threshold degradation, the ability of the baseline reference value to distinguish historically known degradation states is verified. When the composite degradation score of a historically known degradation window is lower than the baseline reference value, the baseline reference value is adjusted downwards. The threshold is stored in the configuration database indexed by product model and is put into use after confirmation by engineers. When the composite degradation score exceeds the threshold, the window is determined to have sub-threshold degradation.

[0064] When historical data is insufficient to support statistical calculations, the initial values ​​of the weighting coefficients for the three characteristic components are estimated a priori based on the electromagnetic coupling physical characteristics of the billet cross-sectional area and aspect ratio. Furthermore, this a priori rule only applies when the billet cross-sectional area is within a specified proportion of the coil's inner diameter cross-sectional area. For cases outside this range, initial values ​​with equal weights for each component are used. The specific method is as follows: Within the range, for blank models with larger cross-sectional areas, the equivalent mutual inductance coupling coefficient between the coil and the blank is larger, the control system is more sensitive to changes in resonant parameters in frequency compensation adjustment, and the prior weighting coefficient of the cumulative change component of frequency control command is relatively high. For billet models with a large length-to-diameter ratio, the non-uniformity of axial current distribution in the billet is more significant, the harmonic characteristics of the resonant circuit change more obviously when the contact resistance increases, and the prior weighting coefficient of the mean difference component of the current harmonics is relatively high. The relative high and low relationships of the aforementioned prior weighting coefficients are confirmed by engineers in the configuration file in the form of normalized values ​​based on the aforementioned qualitative analysis principles, and stored in the configuration database as initial weighting coefficients. When the number of valid historical samples for the same product model reaches the minimum statistical sample number requirement specified in the configuration file, the weighting coefficients are recalculated using the statistical correlation method to replace the prior initial values.

[0065] Specifically, in the initial stage of knowledge graph construction, when historical data is insufficient to support statistical correlation calculations, the weighting coefficients are estimated a priori based on the electromagnetic coupling physical characteristics of the billet cross-sectional area and aspect ratio. This a priori rule only applies to cases where the billet cross-sectional area falls within a specified ratio range of the coil's inner diameter cross-sectional area. This ratio range is confirmed by engineers based on coil design parameters and billet specifications, using historical electromagnetic simulation data or factory test records, and then stored in a configuration file. Models exceeding this range use initial values ​​with equal weights for the three components.

[0066] Within the applicable scope: For billet models with larger cross-sectional areas, the equivalent mutual inductance coupling coefficient between the coil and the billet is greater, the frequency compensation adjustment of the control system is more sensitive to changes in resonance parameters, and the prior weighting coefficient of the cumulative change component of the frequency control command is relatively higher; For billet models with larger length-to-diameter ratios, the axial current distribution non-uniformity of the billet is more significant, the harmonic characteristics change more obviously when the contact resistance increases, and the prior weighting coefficient of the mean difference component of the current harmonics is relatively higher. The prior weighting coefficients are confirmed by engineers in the configuration file in normalized numerical form according to the above qualitative analysis principles and then stored in the configuration database. When the number of valid historical samples for the same product model reaches the minimum statistical sample number requirement specified in the configuration file, the weighting coefficients are recalculated using statistical correlation methods to replace the prior initial values.

[0067] After the entity node without log potential root causes is instantiated, cross-window degradation trend extraction is performed on the degradation feature record sequences corresponding to multiple consecutive shift time windows of the same device and product model, and directed degradation trajectory time-series edges are constructed between the corresponding entity nodes. The specific steps include: For the degradation feature record sequence grouped by device identifier and product model in the feature database and arranged by window start timestamp, the direction and amount of change of each feature component between adjacent window pairs are calculated sequentially. Within the specified number of consecutive windows, when the total amount of change of a certain feature component is not less than the specified proportion of the width of the corresponding subthreshold degradation judgment interval, and the proportion of the number of window pairs increasing in the same direction to the total number of consecutive window pairs is not less than the monotonicity proportion requirement specified in the configuration file, it is determined that the feature component has a monotonic degradation trend in the consecutive window sequence. The minimum value of the number of consecutive windows, the proportion requirement of change, and the monotonicity proportion requirement are all stored in the configuration database with the product model as the index, and are determined by engineers based on the historical normal production fluctuation level of the same model. For a continuous window sequence that is determined to have a monotonic degradation trend, a directed degradation trajectory time-series edge is constructed between the corresponding adjacent unlogged potential root cause entity nodes. The direction of the edge is from the entity node with an earlier time to the entity node with a later time. The attributes of the edge include the change between the corresponding feature component identifier and the adjacent window pair. Interruptions in continuous window sequences are handled according to the following rules: For shutdown interruptions caused by equipment power failure or disconnection of the main circuit, the construction of the degradation trajectory timing edge is cut off at the interruption point, and statistics are restarted from the starting point of a new continuous window sequence after the interruption; For production interruptions caused by equipment being powered on but production being stopped, degradation trend statistics are allowed to continue across the interruption point as long as the interruption time interval does not exceed the maximum allowable interruption duration specified in the configuration file. If the interruption time exceeds this duration, it is treated as a shutdown interruption.

[0068] Specifically, after the unlogged potential root cause entity node is instantiated, the degraded feature record sequence in the feature database, grouped by device identifier and product model and arranged by window start timestamp, is used to calculate the direction and amount of change of each feature component between adjacent window pairs.

[0069] The above-mentioned judgment parameters are obtained as follows: A minimum number of consecutive windows is specified, taken as the upper bound of the distribution of the number of consecutive windows in the same direction during historical normal fluctuations, ensuring that the normal fluctuation sequence does not meet the minimum window number requirement; the change ratio requirement is taken as the upper bound of the distribution of the ratio of the total change to the width of the judgment interval in historical normal fluctuations; the monotonicity ratio requirement is taken as the upper bound of the distribution of the proportion of windows in the same direction in historical normal fluctuation sequences, and adjusted upwards within the correction range specified in the configuration file. All of the above parameters are stored in the configuration database after being confirmed by engineers using the product model as an index.

[0070] Within the specified number of consecutive windows, if the total change of a certain feature component is not less than a specified proportion of the width of the corresponding subthreshold degradation judgment interval, and the proportion of the number of window pairs increasing in the same direction to the total number of consecutive window pairs is not less than the monotonicity proportion requirement specified in the configuration file, then the feature component is determined to have a monotonically degrading trend. For consecutive window sequences determined to have a monotonically degrading trend, a directed degradation trajectory time-series edge is constructed between corresponding adjacent unlogged potential root cause entity nodes, with the direction pointing from the earlier node to the later node. The edge attributes include the change between the corresponding feature component identifier and the adjacent window pair.

[0071] The interruption handling rules for continuous window sequences are as follows: For shutdown interruptions caused by equipment power failure or disconnection of the main circuit, the construction of the degradation trajectory timing edge is cut off at the interruption point, and the statistics are restarted from the starting point of the new continuous window sequence after the interruption; For production interruptions caused by equipment being powered on but production stopping, the degradation trend statistics are allowed to continue across the interruption point as long as the interruption time interval does not exceed the maximum allowable interruption duration specified in the configuration file; The maximum allowable interruption duration is determined based on the statistical data of the equipment resonance parameter drift during historical shutdowns but power-on periods, and the lower bound of the duration distribution at which the drift begins to significantly exceed the normal operation fluctuation range is taken, which is confirmed by the engineers and stored in the configuration file; if this duration is exceeded, it is treated as a shutdown interruption.

[0072] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a fault log knowledge graph for complete sets of electrical equipment, characterized in that, include: Driven by the ontological attribute mode and with the product model switching time as the window boundary, the subthreshold degradation feature record of the resonant circuit is extracted from the induction heating operation parameter sequence. Specifically, this includes: collecting the inverter operating frequency control command sequence, the reactive power compensation circuit switching action record sequence, and the inverter output side current harmonic analysis result sequence from the induction heating control system; defining window boundaries based on the product model switching log, and dividing the continuous production cycle of the same model into time windows by shift; Within each time window, with the window start value as the reference, the cumulative change of the frequency control command within the window, the deviation between the current window reactive power compensation switching frequency and the historical reference switching frequency of the same model, and the difference between the current window current harmonic analysis mean and the historical reference mean of the same model are calculated to form a combination of degradation characteristic components. The degradation feature components are combined with the start and end timestamps of the corresponding window and the product model identifier and encapsulated into a complete degradation feature record; Based on the order of billet parts exiting the furnace, the degradation feature records are aligned with the eccentric load features of the downstream forging process at the part-level time sequence. After correction for the effects of transmission and cooling, cross-process causal evidence records are generated. A semantically symmetric ontology structure is constructed, defining logged fault entity classes and non-logged potential root cause entity classes as mutually exclusive subclasses with semantically symmetric attribute patterns. Non-logged potential root cause entity nodes are instantiated based on degradation feature records. Specifically, this includes: traversing degradation feature records in the feature database and filtering records that simultaneously meet the following conditions: the combination of degradation feature components satisfies the subthreshold degradation judgment condition, and there is at least one downstream forging process fault log record within the corresponding shift time window; creating a non-logged potential root cause entity instance for each degradation feature record that meets the conditions, and sequentially filling the corresponding ontology attributes with the degradation window start and end timestamps, the degradation severity description string generated based on the feature component combination values, the equipment identifier, the product model, and the degradation feature vector value; registering the entity instance with completed attribute filling to the knowledge graph node set, and establishing a mapping relationship between the instance identifier and the feature database record index; Based on cross-process causal evidence records, directed cross-process causal relationship edges are constructed between unlogged potential root cause entity nodes and process-sensitive logged fault entity nodes; classification topology integrity verification is performed on process-sensitive fault entities, and an iterative update process is triggered for isolated symptom nodes to complete the knowledge graph closure construction.

2. The method for constructing a fault log knowledge graph for electrical complete sets of equipment according to claim 1, characterized in that: The steps for generating cross-process causal evidence records include: A pre-established correlation between product models and billet specifications is established, including cross-sectional area and length-to-diameter ratio. Based on historical thermal imaging measurement data of each billet model, the reference component of eccentric load caused by furnace cooling under different transmission delay conditions is calibrated. Using the billet part number as the primary key, the degradation feature record of the shift time window to which the part belongs is extracted from the feature database, and the eccentric load feature value of the part is extracted. The cooling correction reference value is retrieved using the product model and the measured transmission delay as a joint index. The eccentric load feature value is subtracted from the reference value to obtain the corrected eccentric load feature component. When the corrected eccentric load characteristic component exceeds the judgment threshold and a fault log record exists for the item, a causal evidence record is generated.

3. The method for constructing a fault log knowledge graph for electrical complete sets of equipment according to claim 1, characterized in that: Constructing a semantically symmetric ontology structure specifically includes: In the knowledge graph ontology file, under the top-level fault-related entity parent class, logged fault entity subclasses and non-logged potential root cause entity subclasses are created through the subclass axiom, and the two are declared to be mutually exclusive through the disjoint class axiom. Define fault timestamp attribute, fault severity level attribute, device identifier attribute, product model attribute, and log text attribute for the fault entity subclass with logs; define semantically symmetrical attributes for the potential root cause entity subclass without logs: the degradation window start and end timestamp attribute is symmetrical to the fault timestamp attribute, the degradation severity description attribute is symmetrical to the fault severity level attribute, the device identifier attribute and product model attribute are completely isomorphic to the fault entity subclass with logs, and the degradation feature vector attribute stores the combined value of feature components and is symmetrical to the log text attribute.

4. The method for constructing a fault log knowledge graph for electrical complete sets of equipment according to claim 1, characterized in that: The determination of process-sensitive log faulty entity nodes and the construction of cross-process causal relationship edges specifically include: A lookup table of fault codes and fault categories is pre-established. Overload, forging dimensional deviation, and impact load abnormality are marked as process-sensitive, while electrical short circuit and sensor disconnection are marked as equipment body type. During the instantiation of log fault entities, the fault category attribute is filled into the log fault entity instance according to the lookup table. Based on the causal evidence records in the evidence database, a directed cross-process causal relationship edge is created between the process-sensitive logged fault entity instance associated with each record and the corresponding unlogged potential root cause entity instance.

5. The method for constructing a fault log knowledge graph for electrical complete sets of equipment according to claim 1, characterized in that: Perform classification topology integrity verification and trigger iterative update process for process-sensitive fault entities, specifically including: For all fault category attribute nodes with log faults in the knowledge graph, count the number of isolated nodes that do not have any incoming cross-process causal relationship edges, and calculate the proportion of isolated nodes to the total number of process-sensitive fault entity nodes. When the proportion of isolated nodes exceeds the preset integrity threshold, a root cause missing alarm report is generated, which includes a list of isolated nodes and the corresponding item information and shift time window identifier for each node, and is submitted for review. The audit conclusions are written back to the knowledge graph through the ontology extension interface, triggering the generation of supplementary evidence records for the affected items and the construction of corresponding cross-process causal relationship edges; After the supplementary construction is completed, the full classification topology integrity verification is re-executed until the proportion of isolated nodes meets the preset integrity threshold, thus completing the iterative closure of the knowledge graph.

Images