Method and apparatus for providing inverter fault solutions based on knowledge graphs
By using a knowledge graph-based approach, real-time operating data of the frequency converter and a pre-built knowledge graph are used to filter and rank frequency converter fault solutions, which solves the problems of low diagnostic accuracy and efficiency in existing technologies and achieves accurate recommendations and rapid response.
Patent Information
- Application Number
- CN202611205184.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-10
- Publication Date
- 2026-11-17
AI Technical Summary
In the current technology, the diagnosis of frequency converter faults relies on expert experience and paper manuals, resulting in low diagnostic accuracy and efficiency.
A knowledge graph-based approach is adopted. By collecting real-time operating data of the frequency converter, candidate relationship edges associated with fault information are selected from the pre-constructed knowledge graph. The solution in the recommended set is determined by matching the real-time operating data with the constraint interval attributes of the candidate relationship edges, and then sorted and output by weight attributes and urgency.
It enables precise recommendations for inverter fault solutions, improves diagnostic accuracy and efficiency, ensures rapid response in emergency fault scenarios, and avoids the shortcomings of relying on expert experience and paper manuals.
Smart Images

Figure CN123117300A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of inverter fault resolution, and in particular to a method and apparatus for outputting inverter fault solutions based on knowledge graphs. Background Technology
[0002] In industrial automation systems, frequency converters are core control devices, and their fault handling efficiency directly affects the normal operation of the production line. Currently, frequency converter fault diagnosis and solutions rely on expert experience and paper manuals, resulting in low accuracy and efficiency in frequency converter fault diagnosis. Summary of the Invention
[0003] This application provides a method and apparatus for outputting inverter fault solutions based on knowledge graphs, in order to solve the problem that the existing technology relies on expert experience and paper manuals for inverter fault diagnosis, resulting in low accuracy and efficiency in inverter fault diagnosis.
[0004] Firstly, this application provides a method for outputting inverter fault solutions based on knowledge graphs, including: In response to the fault information reported by the frequency converter, the real-time operating condition data of the frequency converter is collected; Based on the fault information, candidate relationship edges associated with the fault information are filtered from a pre-constructed knowledge graph. The real-time operating condition data is then matched with the constraint interval attributes of each candidate relationship edge to add the solution entity corresponding to the successfully matched candidate relationship edge to the recommendation set. The knowledge graph includes multiple types of entities and multiple relationship edges connecting the entities. Each relationship edge is associated with a fault entity, an operating condition parameter entity, an environmental parameter entity, and a solution entity, and each relationship edge is attached with constraint interval attributes and weight attributes. The comprehensive score of each solution in the recommendation set is determined based on the weight attribute of each candidate relation edge, and the solutions in the recommendation set are sorted from high to low according to the comprehensive score. The solution to be output is determined from the sorted solutions according to the urgency of the fault information.
[0005] Secondly, this application provides a device for outputting inverter fault solutions based on knowledge graphs, comprising: The data acquisition module is used to collect real-time operating data of the frequency converter in response to the fault information reported by the frequency converter. The first processing module is used to filter candidate relationship edges associated with the fault information from a pre-constructed knowledge graph based on the fault information, and match the real-time operating condition data with the constraint interval attributes of each candidate relationship edge, so as to add the solution entity corresponding to the successfully matched candidate relationship edge to the recommendation set; wherein, the knowledge graph includes multiple types of entities and multiple relationship edges connecting the entities, each relationship edge is associated with a fault entity, an operating condition parameter entity, an environmental parameter entity and a solution entity, and each relationship edge is attached with a constraint interval attribute and a weight attribute; The second processing module is used to determine the comprehensive score of each solution in the recommendation set according to the weight attribute of each candidate relation edge, sort the solutions in the recommendation set from high to low according to the comprehensive score, and determine the solution to be output from the sorted solutions according to the urgency of the fault information.
[0006] Thirdly, this application provides an electronic device, comprising: at least one communication interface; at least one bus connected to the at least one communication interface; at least one processor connected to the at least one bus; and at least one memory connected to the at least one bus, wherein the processor is configured to execute the method for outputting a frequency converter fault solution based on a knowledge graph as described in the first aspect of this application.
[0007] Fourthly, this application also provides a computer storage medium storing computer-executable instructions for executing the method for outputting inverter fault solutions based on knowledge graphs as described in the first aspect of this application.
[0008] Compared with the prior art, the above-mentioned technical solution provided in this application has the following advantages: In response to the fault information reported by the frequency converter, the real-time operating condition data of the frequency converter is collected. Then, based on the fault information, candidate relationship edges associated with the fault information are filtered from the pre-constructed knowledge graph. The real-time operating condition data is matched with the constraint interval attributes of each candidate relationship edge to add the solution entity corresponding to the successfully matched candidate relationship edge to the recommendation set. Finally, the comprehensive score of each solution in the recommendation set is determined according to the weight attribute of each candidate relationship edge. The solutions in the recommendation set are sorted from high to low according to the comprehensive score. The required solution is output by combining the urgency of the fault information and the sorted solutions. As can be seen, in this embodiment, the structured knowledge representation of the knowledge graph and the conditional matching retrieval based on real-time operating data enable accurate recommendations for inverter fault solutions. This avoids the problem of low accuracy and efficiency in inverter fault diagnosis caused by relying on expert experience and paper manuals in the prior art. In addition, the credibility of each solution is quantitatively compared and ranked by weighted attributes and a comprehensive scoring mechanism. The output strategy of urgency level classification ensures rapid response in emergency fault scenarios. Attached Figure Description The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.
[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0010] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.
[0011] Figure 1 A flowchart illustrating a method for providing a fault solution for a frequency converter based on a knowledge graph, as provided in this application embodiment; Figure 2 A schematic diagram of a quadruple conditional knowledge graph construction system for inverter fault diagnosis provided in an embodiment of this application; Figure 3 This is a schematic diagram of the quadruple conditional knowledge graph structure provided in an embodiment of this application; Figure 4A flowchart illustrating the conditional search process based on real-time operating conditions provided in this application embodiment; Figure 5 This application provides a flowchart of edge computing and cloud-based collaborative inference in its embodiments. Figure 6 A schematic diagram of the recommended output interface provided for embodiments of this application; Figure 7 A schematic diagram of the device structure for providing a knowledge graph-based frequency converter fault solution in an embodiment of this application; Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation
[0012] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0013] The following disclosure provides numerous different embodiments or examples for implementing various structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. These are merely examples and are not intended to limit the scope of the invention. Furthermore, reference numerals and / or letters may be repeated in different examples. Such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.
[0014] To address the problem of low accuracy and efficiency in inverter fault diagnosis due to reliance on expert experience and paper manuals in existing technologies, this application provides a method for outputting inverter fault solutions based on knowledge graphs, such as... Figure 1 As shown, the steps of this method include: Step 101: In response to the fault information reported by the frequency converter, collect the real-time operating condition data of the frequency converter; In this embodiment, the fault information reported by the frequency converter refers to the fault notification message sent by the frequency converter to the upper-level system through its communication interface when it detects an abnormal operating state during operation. This fault information includes at least a fault code, such as OC (overcurrent fault), OV (overvoltage fault), and OH (overheating fault). The fault code serves as the retrieval key for subsequent knowledge graph searches.
[0015] Furthermore, collecting real-time operating data of the frequency converter refers to obtaining the current operating status parameters and environmental parameters from the frequency converter and its operating environment at the moment a fault is reported, via sensor networks, data buses, or data interfaces. Specific content of real-time operating data may include load rate (e.g., the percentage of the current load at rated load), speed, running time, radiator temperature, bus voltage, ambient temperature, and ambient humidity. Real-time processing is required because frequency converter fault diagnosis is highly dependent on the on-site operating conditions at the moment the fault occurs; outdated data cannot accurately reflect the true conditions at the time of the fault.
[0016] For example, a frequency converter triggers an overcurrent fault due to an abnormal increase in current during operation, reporting a fault code OC to the system. The system then uses sensors to read the current load rate as 92%, the radiator temperature as 68℃, and the ambient temperature as 34℃. This data reflects the actual operating status of the frequency converter at the moment the fault occurred and is crucial for determining the recommended solution. Therefore, automatically collecting real-time data from the field at the first moment of fault triggering ensures that the operating condition information used for subsequent matching is consistent with the actual conditions at the time of the fault, avoiding inaccurate recommendations due to the use of static or outdated data.
[0017] Step 102: Based on the fault information, candidate relationship edges associated with the fault information are selected from the pre-constructed knowledge graph. The real-time operating condition data is then matched with the constraint interval attributes of each candidate relationship edge to add the solution entity corresponding to the successfully matched candidate relationship edge to the recommendation set. The knowledge graph includes multiple types of entities and multiple relationship edges connecting entities. Each relationship edge is associated with a fault entity, an operating condition parameter entity, an environmental parameter entity, and a solution entity. Each relationship edge is also associated with a constraint interval attribute and a weight attribute. In this embodiment, the knowledge graph is a structured knowledge representation model, whose basic components include entities and relational edges. Further, an entity refers to a core conceptual object in the field of inverter fault diagnosis, including at least fault entities (such as overcurrent faults), operating condition parameter entities (such as load rate), environmental parameter entities (such as ambient temperature), and solution entities (such as replacing IGBT modules). A relational edge is a directed knowledge connection linking different entities; each relational edge represents a complete knowledge unit. Furthermore, "based on fault information" means using the fault code in the fault information as the retrieval key. Since each relational edge in the knowledge graph is associated with a fault entity, using the fault code as the key allows for rapid location of all relational edges related to that fault type from the entire graph. Additionally, filtering candidate relational edges associated with fault information means extracting all relational edges from the knowledge graph that originate from that fault entity. These candidate relational edges represent all knowledge entries that may be applicable to that fault type, but which ones are truly applicable to the current site needs further determination through operating condition matching.
[0018] For example, using fault code OC as the key, two relational edges originating from the overcurrent fault entity are found in the knowledge graph: edge A relates to light load conditions and solutions for increasing acceleration time, and edge B relates to heavy load conditions and solutions for replacing the IGBT module; both edges are included in the candidate relational edge set. It is evident that by using the fault code as the key for initial screening in the knowledge graph, the search scope is narrowed from the entire graph to a knowledge set only related to this fault type, significantly reducing the computational load of subsequent matching and improving search efficiency.
[0019] In this embodiment, the constraint interval attribute is an attribute field attached to the relation edge, used to limit the applicable operating conditions of the knowledge represented by the relation edge. Each relation edge has at least two constraint intervals: operating condition constraint interval (such as the applicable interval of load rate) and environmental constraint interval (such as the applicable interval of ambient temperature). Each constraint interval is defined by a numerical upper limit and a numerical lower limit. Therefore, matching the real-time operating condition data with the constraint interval attributes of each candidate relation edge means comparing the collected real-time operating condition data (including operating condition parameter values and environmental parameter values) one by one with the constraint interval of each candidate relation edge. Furthermore, adding the solution entity corresponding to the successfully matched candidate relation edge to the recommendation set means that only the solutions pointed to by relation edges whose constraint intervals contain the current real-time data are included in the subsequent scoring and ranking process. This means that any knowledge item whose constraints do not match the actual situation on site, even if its fault code matches, is excluded from the recommendation range and will not interfere with the user.
[0020] For example, in the overcurrent fault example, the operating condition constraint range of edge A is the load rate [0%, 30%], and the operating condition constraint range of edge B is the load rate [80%, 100%]. When the current load rate is collected as 92%, edge A does not meet the matching condition (92% does not belong to [0%, 30%]), while edge B meets the matching condition (92% belongs to [80%, 100%]). If the ambient temperature of 34℃ also falls within the environmental constraint range of edge B [40℃, 60℃] (it should be noted that 34℃ does not fall within the range [40℃, 60℃]. This example is used in the conditional retrieval step to illustrate the situation where edge B is excluded due to unmet environmental conditions; this will not be elaborated upon here, only using the load rate condition to illustrate the working principle of the matching mechanism), then the relationship edge is successfully matched, and its corresponding solution of replacing the IGBT module is added to the recommendation set. It can be seen that by accurately comparing the real-time operating condition data with the constraint range of the relationship edge, it ensures that only knowledge items specifically applicable to the current field conditions are recommended to the user. This matching method based on numerical range inclusion judgment has the advantages of high determinism, simple calculation and strong interpretability, and is particularly suitable for efficient execution on edge computing devices.
[0021] Furthermore, in this application, it should be noted that "multi-type entities" refers to the classification of nodes in the knowledge graph according to the types of objective objects they represent, with different types of entities storing different categories of information. Fault entities represent various types of faults that may occur in the frequency converter, such as overcurrent, overvoltage, overheating, grounding, and communication anomalies, with each fault entity having a fault severity level attribute. Operating condition parameter entities represent key operating parameters of the frequency converter during operation, including but not limited to load rate, speed, running time, radiator temperature, and bus voltage, with each operating condition parameter entity having a normal range threshold attribute. Environmental parameter entities represent the state parameters of the external environment in which the frequency converter operates, including but not limited to ambient temperature, ambient humidity, altitude, and electromagnetic interference level. Solution entities represent specific handling measures that can be taken under specific conditions for a specific fault, including but not limited to parameter adjustment steps, component replacement procedures, software upgrade methods, and hardware reset operations, with each solution entity having an operation complexity attribute and an expected time consumption attribute. Relationship edges are directed connections linking the above-mentioned entities. In this scheme, a relation edge simultaneously associates entities across four dimensions: it originates from a fault entity, associates with a condition parameter entity, an environmental parameter entity, and points to a solution entity. This structure, which simultaneously associates four entities, is called a quadruple structure, mathematically meaning that under specific operating conditions and environmental conditions, a certain fault can be handled using a specific solution. The constraint interval attribute is one of the core attributes of a relation edge. As mentioned earlier, it is used to limit the applicable conditions of the relation edge, including the operating condition constraint interval and the environmental constraint interval. The weight attribute is another core attribute of a relation edge, used to quantify the confidence or reliability of the knowledge represented by the relation edge. The weight attribute can include static weights (initial confidence preset by experts during knowledge construction) and dynamic weights (current confidence continuously adjusted based on actual user feedback).
[0022] For example, in a knowledge graph, a complete relational edge record is as follows: the fault entity is overcurrent, the operating parameter entity is load rate, the operating constraint interval is [80%, 100%], the environmental parameter entity is ambient temperature, the environmental constraint interval is [40℃, 60℃], the solution entity is replacing the IGBT module, the static weight is 0.92, and the dynamic weight is 0.92. This record fully expresses the following knowledge: when an overcurrent fault occurs, if the current load rate is between 80% and 100% and the ambient temperature is between 40℃ and 60℃, then replacing the IGBT module is recommended. The initial confidence level of this knowledge is 0.92. Through the data structure of quadruples plus constraint intervals, each piece of knowledge in the knowledge graph fully retains its applicable preconditions, avoiding the misuse of knowledge caused by the loss of conditional information in traditional knowledge representation. At the same time, the weight attribute provides a quantitative basis for subsequent solution ranking and also provides an adjustable dimension for knowledge evolution based on user feedback.
[0023] As can be seen, the modeling approach using a quadruple (fault entity, operating condition entity, environment entity, solution entity) plus explicit constraint attributes adds applicable operating condition and environment ranges to each matching relationship. Compared to the existing triple (fault-relationship-solution) knowledge graph, which has strong conditional dependencies—for example, "overcurrent" faults are often due to parameter drift under light load (load rate <30%), which can be resolved by adjusting the acceleration time; under heavy load (load rate >80%), they are often due to hardware damage, requiring replacement of the power module—the quadruple knowledge graph in this application can completely preserve the multivariate constraints in fault diagnosis, avoiding information distortion. For inverter fault scenarios, the accuracy of conditional retrieval can be significantly improved, while the retrieval calculation only involves numerical range judgment, and the edge reasoning delay can be controlled within a short time.
[0024] Step 103: Determine the comprehensive score of each solution in the recommendation set based on the weight attributes of each candidate relation edge, sort the solutions in the recommendation set from high to low according to the comprehensive score, and determine the solution to be output from the sorted solutions according to the urgency of the fault information.
[0025] To address this, since the recommendation set may contain multiple solutions (for example, in relaxed matching mode, there may be multiple relation edges with constraint intervals close to the current data), a unified quantitative metric is needed to compare the merits of each solution in order to provide users with a ranked recommendation list. The weight attribute reflects the credibility of the knowledge represented by each relation edge. The higher the weight value, the more reliable the knowledge. The comprehensive score is the final evaluation score calculated based on the weight attribute according to preset rules. In this method, the comprehensive score can be the sum of the static and dynamic weights multiplied by their corresponding weighting coefficients. When multiple relation edges point to the same solution, the weights can be superimposed during the calculation process, or the highest value can be taken. The specific calculation method does not affect the core function of this step, namely, to quantitatively evaluate each solution in the recommendation set. Therefore, by quantifying the credibility of each knowledge item into a comprehensive score, an objective numerical basis is provided for comparing and ranking solutions, ensuring that the final recommendation output is not arbitrary or based on simple rules, but rather an optimal selection based on knowledge confidence.
[0026] Furthermore, sorting by the comprehensive score from highest to lowest means arranging all solutions in the recommendation set in descending order of their comprehensive scores. In the sorted list, the higher the ranking of a solution, the higher its comprehensive score, indicating a higher confidence level in the system's recommendation of that solution, and thus it should be given priority. Providing a ranked output of solutions, rather than just outputting a single best solution, has significant technical implications. On one hand, when the highest-scoring solution lacks sufficient confidence, users can refer to the second or third-ranked solutions; on the other hand, in actual operation, maintenance personnel sometimes need to choose the most feasible option from multiple feasible solutions based on the specific on-site conditions, and providing a ranked list can meet this need for flexibility.
[0027] Furthermore, the urgency level of fault information refers to the severity level of the fault itself. This level can be pre-stored in the fault entity or predefined in the system according to the fault type. For example, grounding faults and bus overvoltages are usually emergency faults requiring immediate handling; overcurrent faults may be considered routine faults under light load conditions but may escalate to emergency faults under heavy load conditions. Combining the urgency level of fault information means that the output strategy is not fixed but dynamically adjusted according to the severity level of the fault. Outputting the required solution refers to deciding whether to output one or multiple solutions from the ranked list of recommended solutions, based on the urgency level. Specifically, for emergency faults, due to extremely limited handling time, outputting multiple solutions would increase the operator's decision-making burden and delay handling; therefore, only the solution with the highest overall score is output. For routine faults, operators have ample time to make judgments, and a complete ranked list of solutions can be output for their reference and selection.
[0028] As can be seen, through the above steps 101 to 103, in response to the fault information reported by the frequency converter, the real-time operating condition data of the frequency converter is collected. Then, based on the fault information, candidate relationship edges associated with the fault information are filtered from the pre-constructed knowledge graph. The real-time operating condition data is matched with the constraint interval attributes of each candidate relationship edge to add the solution entity corresponding to the successfully matched candidate relationship edge to the recommendation set. Finally, the comprehensive score of each solution in the recommendation set is determined according to the weight attribute of each candidate relationship edge. The solutions in the recommendation set are sorted from high to low according to the comprehensive score. The required solution is output by combining the urgency of the fault information and the sorted solutions. As can be seen, in this embodiment, the structured knowledge representation of the knowledge graph and the conditional matching retrieval based on real-time operating data enable accurate recommendations for inverter fault solutions. This avoids the problem of low accuracy and efficiency in inverter fault diagnosis caused by relying on expert experience and paper manuals in the prior art. In addition, the credibility of each solution is quantitatively compared and ranked by weighted attributes and a comprehensive scoring mechanism. The output strategy of urgency level classification ensures rapid response in emergency fault scenarios.
[0029] In an optional embodiment of this application, the method of matching real-time operating data with the constraint interval attributes of each candidate relationship edge in step 102 above, so as to add the solution entity corresponding to the successfully matched candidate relationship edge to the recommendation set, may further include: Step 11: Match the operating parameters in the real-time operating data with the operating constraint interval in the constraint interval attribute, and match the environmental parameters in the real-time operating data with the environmental constraint interval in the constraint interval attribute. In this embodiment, operating condition parameters refer to data characterizing the inverter's own operating status, including but not limited to load rate, speed, running time, radiator temperature, and bus voltage. These parameters reflect the inverter's operating status and are the core basis for determining the cause of a fault and selecting a solution. Furthermore, operating condition constraint intervals are attribute fields pre-stored on relational edges, used to limit the range of operating condition parameters within which the knowledge represented by the relational edge is valid. Each operating condition constraint interval is defined by an upper limit and a lower limit.
[0030] Furthermore, environmental parameters refer to data characterizing the external environmental conditions of the frequency converter, including but not limited to ambient temperature, humidity, altitude, and electromagnetic interference levels. Unlike operating condition parameters, environmental parameters reflect the impact of external conditions on the operation of the frequency converter, rather than the operating state of the frequency converter itself. In frequency converter fault diagnosis, environmental conditions (especially temperature) have a significant impact on the causes of faults; for example, the causes of the same fault may be completely different in high-temperature environments and normal-temperature environments. Environmental constraint intervals are another attribute field pre-stored on relation edges, used to limit the range of environmental parameters within which the knowledge represented by the relation edge is valid. Each environmental constraint interval is also defined by an upper limit and a lower limit.
[0031] Step 12: Add the solution entities corresponding to the successfully matched candidate relationship edges to the recommendation set.
[0032] As can be seen, in this embodiment, by splitting real-time operating data into two dimensions—operating parameters and environmental parameters—and matching them with two types of constraint intervals on the relation edges, accurate judgment of the applicability of knowledge is achieved. Operating parameters reflect the equipment's own state, while environmental parameters reflect external conditions; together, these two types of parameters constitute the complete context for fault diagnosis.
[0033] In an optional embodiment of this application, the method of adding the solution entity corresponding to the successfully matched candidate relationship edge to the recommendation set in step 12 above may further include: Step 21: If the operating parameters fall within the operating condition constraint range and the environmental parameters fall within the environmental constraint range, then the real-time operating data and the candidate relationship edge are successfully matched. As can be seen, step 21 above defines the criteria for successful exact matching. Exact matching uses AND logic, meaning that a candidate relation edge is considered a successful match only when the current values of both the operating condition parameter and the environmental parameter simultaneously meet the upper and lower limits of their respective constraints. Both conditions are indispensable. In the case of a successful exact match, the solution entity pointed to by the candidate relation edge is added to the recommendation set. For example, if the current load rate is 25% and the ambient temperature is 28℃, the operating condition constraint range for relation edge A is [0%, 30%] and the environmental constraint range is [-10℃, 40℃]. Since both conditions are met simultaneously, an exact match is successful, and the solution for that relation edge is directly added to the recommendation set. This is the normal matching path.
[0034] Step 22: If the operating parameters do not fall within the operating constraint interval or the environmental parameters do not fall within the environmental constraint interval, determine the first difference between the operating parameters and the midpoint of the operating constraint interval of each candidate relation edge, and determine the second difference between the environmental parameters and the midpoint of the environmental constraint interval of each candidate relation edge. As can be seen, if any of the operating parameters or environmental parameters does not meet its corresponding constraint interval, the exact match is considered to have failed, triggering a relaxed match. This means that in the event of an exact match failure, the system will not directly abandon the candidate relationship edge, but will initiate a substitute matching mechanism to attempt to evaluate how close the relationship edge is to the current actual operating conditions. The purpose of this is to prevent situations where there are no matching results due to overly strict constraint interval settings.
[0035] For example, in an overcurrent fault scenario, the system collects data showing a current load rate of 90% and an ambient temperature of 34℃. The load condition constraint range for candidate relation edge B is [80%, 100%], and the environmental constraint range is [40℃, 60℃]. The system determines that the 90% load rate falls within [80%, 100%], thus satisfying the load condition; however, the 34℃ ambient temperature does not fall within [40℃, 60℃], therefore, the environmental condition is not satisfied. Because the environmental condition is not met, exact matching fails, and the system triggers relaxed matching.
[0036] Furthermore, the midpoint of the operating condition constraint interval refers to the arithmetic mean of the upper and lower limits of the operating condition constraint interval. For example, when the constraint interval is [80%, 100%], the midpoint is (80% + 100%) ÷ 2 = 90%; when the constraint interval is [0%, 30%], the midpoint is (0% + 30%) ÷ 2 = 15%. The midpoint of the interval represents the optimal or central value of the operating condition parameter to which this knowledge applies. The first difference refers to the absolute difference between the current real-time operating condition parameter value and the midpoint of the operating condition constraint interval. This difference reflects the degree of deviation between the current operating condition and the optimal operating condition to which this knowledge applies. The smaller the difference, the closer the current operating condition is to the optimal condition designed for this relation edge. For example, the operating condition constraint interval of candidate relation edge B is [80%, 100%], and the midpoint is 90%. The current load rate is 90%, and the first difference is |90% - 90%| = 0. The load condition constraint interval for candidate relation edge C is [40%, 60%], with a midpoint of 50%. The current load rate is 90%, and the first difference is |90%-50%|=40%. The first difference of 0 is much less than 40, indicating that relation edge B is closer to the current actual load condition than relation edge C in terms of load condition dimension.
[0037] Furthermore, the midpoint of the environmental constraint interval refers to the arithmetic mean of the upper and lower limits of the environmental constraint interval. Its calculation method is the same as that for the midpoint of the operating condition constraint interval. The second difference refers to the absolute difference between the current real-time environmental parameter value and the midpoint of the environmental constraint interval. This difference reflects the degree of deviation between the current environmental conditions and the optimal applicable environmental conditions for this knowledge. Similar to the first difference, the smaller the second difference, the closer the current environment is to the optimal environmental conditions designed for this relation edge. For example, the environmental constraint interval of candidate relation edge B is [40℃, 60℃], with a midpoint of 50℃. The current environmental temperature is 34℃, and the second difference is |34℃-50℃|=16℃. The environmental constraint interval of candidate relation edge D is [20℃, 35℃], with a midpoint of 27.5℃. The current environmental temperature is 34℃, and the second difference is |34℃-27.5℃|=6.5℃. The environmental deviation of relation edge D is less than that of relation edge B.
[0038] Step 23: Determine the overall deviation of each candidate relationship edge from the real-time operating data based on the first difference and the second difference, and determine the candidate relationship edge whose overall deviation is less than the first preset threshold and successfully match the real-time operating data. It should be noted that if there are no candidate relationship edges with a comprehensive deviation less than the first preset threshold, then no recommendation result will be returned.
[0039] In this embodiment, the comprehensive deviation degree is a single quantitative index obtained by comprehensively processing the first difference and the second difference, used to comprehensively measure the overall distance between a candidate relationship edge and the current actual working condition. The specific calculation method for the comprehensive deviation degree can be a weighted sum, simple sum, or sum of squares of the first and second differences, or other forms of comprehensive calculation. The smaller the value of the comprehensive deviation degree, the closer the candidate relationship edge is to the current actual working condition. For example, if the first difference of candidate relationship edge B is 0 and the second difference is 16, the comprehensive deviation degree can be a simple sum of 16, or a weighted sum (e.g., load rate weight 0.7 × 0 + temperature weight 0.3 × 16 = 4.8). Candidate relationship edge D has a large deviation at the midpoint of its working condition constraint interval (first difference 40) and a small environmental deviation (second difference 6.5), so the comprehensive deviation degree is 40 + 6.5 = 46.5, or a weighted sum of 0.7 × 40 + 0.3 × 6.5 = 29.95. After comprehensive comparison, the overall deviation of relation edge B (16 or 4.8) is significantly smaller than that of relation edge D (46.5 or 29.95), and relation edge B is closer to the current working condition overall.
[0040] It should be noted that the larger the first preset threshold, the higher the tolerance for deviation and the more candidate relationship edges are successfully matched, but the accuracy of the recommendation results may decrease; the smaller the threshold, the higher the requirement for precise matching and the more accurate the recommendation results, but there may be a risk of having no recommendations to generate.
[0041] Step 24: Add the solution entities corresponding to the successfully matched candidate relationship edges to the recommendation set.
[0042] As can be seen, steps 21 to 24 above resolve the issue of empty results that may result from exact matching. In real-world industrial scenarios, due to the complex and variable nature of on-site conditions, real-time collected data may not fall precisely within any predefined knowledge constraint range. Without a relaxed matching mechanism, the system would produce no recommendations and lose its practical value. Relaxed matching, by allowing a certain degree of deviation, ensures that the system can provide recommendations in the vast majority of scenarios.
[0043] In an optional embodiment of this application, the method of determining the comprehensive score of each solution in the recommendation set based on the weight attributes of each candidate relation edge involved in step 103 above may further include: Step 31: Multiply the preset first weighting coefficient by the static weight in the weight attribute to obtain the first product result; In this embodiment, the static weight is a confidence value preset by domain experts during the construction of the knowledge graph, used to characterize the credibility of the knowledge represented by the relation edge in the early stages of construction. The static weight is a fixed initial value, and its setting can be based on expert experience, historical statistical data, experimental verification results, etc. For example, a piece of knowledge that has been verified by a large amount of field practice can be assigned a higher static weight (e.g., 0.95), while a piece of knowledge from a single document or case experience is assigned a lower static weight (e.g., 0.60).
[0044] Furthermore, the first weighting coefficient is a preset value used to control the proportion of static weights in the overall score. The specific value of the first weighting coefficient can be set according to actual application needs, such as 0.4, 0.5, or 0.7. The larger the first weighting coefficient, the greater the influence of static weights on the overall score, and the higher the system's reliance on expert-preset knowledge; the smaller the first weighting coefficient, the smaller the influence of static weights on the overall score, and the higher the system's reliance on dynamic weights.
[0045] Step 32: Multiply the preset second weighting coefficient with the dynamic weight in the weight attribute to obtain the second product result; In this embodiment, the dynamic weight is a confidence value that is continuously adjusted based on user feedback during the use of the relation edge. It is used to characterize the current credibility of the knowledge represented by the relation edge after actual use and verification. The initial value of the dynamic weight is equal to the static weight, that is, at the beginning of the knowledge graph construction, the dynamic weight and the static weight are the same. As the system continuously receives valid or invalid feedback from users in actual use, the dynamic weight is incrementally adjusted according to a preset update rule (such as increasing or decreasing by 0.1), gradually deviating from the initial static weight. After sufficient user feedback, the dynamic weight will more realistically reflect the effectiveness of the knowledge in the actual situation than the static weight.
[0046] Furthermore, the second weighting coefficient is a preset value used to control the proportion of dynamic weights in the overall score. The second weighting coefficient, together with the first weighting coefficient, determines the relative importance of static and dynamic weights in the overall score. When the second weighting coefficient is greater than the first weighting coefficient, the dynamic weights have a greater impact on the overall score than the static weights, placing greater emphasis on knowledge validated by practical application feedback; conversely, it places greater emphasis on knowledge preset by experts.
[0047] It should be noted that the first weighting coefficient and the second weighting coefficient are usually complementary, meaning their sum equals 1. This complementarity ensures that the numerical range of the overall score is consistent with the numerical range of the weighting attribute itself, facilitating subsequent confidence level assessment and threshold comparison.
[0048] Step 33: The sum of the first product result and the second product result is used as the comprehensive score.
[0049] In this embodiment, the relative magnitudes of the first weighting coefficient and the second weighting coefficient determine the system's preference for expert prior knowledge and user practical feedback. In practical applications, the relative magnitudes of the two can be dynamically adjusted according to different stages: In the initial deployment phase, due to the limited accumulation of user feedback data and the insufficient validation of the dynamic weights, the first weighting coefficient can be set to be greater than the second weighting coefficient (e.g., the first weighting coefficient is 0.7 and the second weighting coefficient is 0.3), so that the system mainly relies on the initial judgment of experts and avoids the limited user feedback in the early stages from excessively affecting the recommendation results; As the system runs longer and user feedback data gradually accumulates, the dynamic weights can more realistically reflect the effectiveness of knowledge in real-world scenarios, so the first weighting coefficient can be adjusted to be less than the second weighting coefficient (e.g., the first weighting coefficient is 0.4 and the second weighting coefficient is 0.6), so that the system gradually shifts towards relying primarily on on-site feedback.
[0050] For example, using a first weighting coefficient of 0.4 and a second weighting coefficient of 0.6, a certain solution has a static weight of 0.85 and a dynamic weight of 0.75. The overall score is 0.4 × 0.85 + 0.6 × 0.75 = 0.34 + 0.45 = 0.79. This score is lower than the static weight of 0.85 but higher than the dynamic weight of 0.75, reflecting the system's balance between expert presets (static weight) and on-site feedback (dynamic weight).
[0051] Through steps 31 to 33 above, static weights represent the professional judgment of domain experts during knowledge graph construction, while dynamic weights represent the verification results from user feedback in actual use. By weighting and combining the two, the overall score neither excludes the initial judgment of experts nor fails to reflect the actual verification results in practice. Therefore, through a unified and quantifiable scoring framework, and by weighting and combining static and dynamic weights, the credibility of different relational edges and different solutions can be compared on the same scale, providing an objective numerical basis for subsequent ranking output.
[0052] In optional embodiments of this application, the method may further include: Step 41: Receive feedback from the user regarding the sorted solutions in the current output; In this embodiment, feedback information refers to the user's evaluation of the effectiveness of the recommended solution in actual fault handling after using it. The feedback information can take the form of binary options, such as effective or invalid, or it can be a quantitative evaluation using a rating system, such as a five-point rating or a percentage rating. Regardless of the specific form, the core function of the feedback information is to identify whether the adopted recommended solution successfully resolved the fault in practical application.
[0053] For example, the system outputs a sorted list of recommended solutions, with the top recommendation being "replace the IGBT module." After the maintenance personnel followed this solution, the inverter returned to normal operation. The personnel clicked the "Valid" button on the system's recommendation output interface to submit feedback. Upon receiving this feedback, the system triggered a subsequent weight update process. Conversely, if the fault was not resolved after following the recommended solution, the personnel clicked the "Invalid" button. This demonstrates that by setting up a convenient feedback collection entry point on the recommendation output interface, the actual effectiveness evaluation of each recommended knowledge can be obtained with extremely low user operation costs, providing a data foundation for the continuous evolution of the knowledge graph.
[0054] Step 42: Perform incremental updates on the weight attributes corresponding to the successfully matched candidate relation edges based on the feedback information.
[0055] In this context, a successfully matched candidate relation edge refers to a relation edge that, during this round of recommendation, is selected through conditional matching (exact matching or relaxed matching) and ultimately has its solution entity added to the recommendation set. When a user selects a solution from the sorted list of recommended solutions and submits feedback, the system needs to locate the one or more relation edges corresponding to that solution, because the weight attribute is bound to the relation edge, not to the solution entity.
[0056] It should be noted that a solution entity may be associated with the same fault entity through multiple relational edges (e.g., multiple relational edges pointing to the same solution under different operating conditions). When a user submits feedback on a recommended solution, the system needs to locate the specific relational edge that was actually matched in this recommendation, that is, the relational edge whose constraint range includes the current real-time operating condition data. Since the system has already recorded the matching status of each candidate relational edge with the current operating condition during the matching step, it can accurately trace which relational edge was used for this recommendation.
[0057] Furthermore, incremental updates of weight attributes refer to increasing or decreasing the weight value of a relation edge by a certain margin. Incremental updates are a gradual adjustment method; each update changes the weight value by only one step (e.g., increasing or decreasing by 0.1), rather than setting the weight to a fixed value. This gradual adjustment avoids drastic fluctuations in weight due to a single abnormal feedback, making the weight evolution process smoother and more stable. When the dynamic weight of a relation edge changes, its overall score will increase or decrease accordingly in the next retrieval, thus changing its ranking in the recommendation list. This achieves a complete closed loop of feedback → weight adjustment → subsequent recommendation ranking changes.
[0058] As can be seen, since each relation edge corresponds to a specific operating condition constraint interval and environmental constraint interval, updating the relation edge weights only affects the knowledge confidence under that specific condition. For example, if a user submits invalid feedback on increasing acceleration time under light load conditions, only the weight of that solution under light load conditions will decrease; the weight of the same solution corresponding to the same fault under heavy load conditions will remain unaffected. This precise positioning characteristic allows the evolution of the knowledge graph to distinguish the validity of knowledge under different operating conditions, avoiding the global interference caused by the one-size-fits-all updates in existing solutions.
[0059] For example, in a light-load (25% load) overcurrent fault scenario, increasing the acceleration time is recommended via relation edge A (operating range [0%, 30%]). User feedback is ineffective, and the system reduces the dynamic weight of relation edge A. Simultaneously, another relation edge B (operating range [80%, 100%]) also points to the solution of increasing the acceleration time (assuming this solution is also an alternative under heavy load conditions). However, since the recommendation hits relation edge A instead of relation edge B, the weight of relation edge B remains completely unaffected. When the system subsequently encounters a heavy-load overcurrent fault, the recommendation priority of increasing the acceleration time under heavy load conditions is irrelevant to the feedback in the light-load scenario.
[0060] As can be seen, through steps 41 and 42 above, continuous knowledge evolution is achieved. That is, the system is no longer a static knowledge base, but a dynamic system capable of learning from each actual use and continuously optimizing itself. With increasing usage time, the weight of user-verified valid knowledge gradually increases, while the weight of invalid knowledge gradually decreases, continuously improving the system's recommendation quality. Furthermore, the weight updates are precise down to each relation edge, that is, precise to the knowledge granularity of a specific fault corresponding to a specific solution under specific operating conditions. This ensures that updates do not interfere with other irrelevant knowledge in the knowledge graph, achieving a balance between local optimization and global stability.
[0061] Based on this, the method of incrementally updating the weight attributes corresponding to the successfully matched candidate relation edges in step 42 above, according to the feedback information, can further include: Step 51: If the feedback information is valid, increase the dynamic weight in the weight attribute by the first step length value, and compare the increased dynamic weight with the preset second preset threshold, and take the smaller value of the two as the updated dynamic weight. As can be seen, step 51 comprises two sequentially executed operations. First, the current value of the dynamic weight is increased by a first-step length value. This first-step length value is a preset positive value used to control the magnitude of weight increase resulting from each effective feedback. The first-step length value should be set appropriately; too large a value may lead to an excessive impact of a single feedback on the weight, while too small a value may result in slow knowledge convergence. Then, the increased dynamic weight is compared with a preset second threshold, and the smaller of the two values is taken as the updated dynamic weight. This step is essentially an upper limit truncation mechanism. The second preset threshold is the maximum value that the dynamic weight is allowed to reach, usually set to 1.0, because weight values are typically defined within the range of 0 to 1, and exceeding 1.0 loses its significance as a confidence level. By taking the smaller value, it ensures that the updated dynamic weight will never exceed the second preset threshold.
[0062] For example, the current dynamic weight of a relationship edge is 0.92, the first step length is 0.1, and the second preset threshold is 1.0. When the user reports "valid," first calculate 0.92 + 0.1 = 1.02, then compare 1.02 with 1.0, and take the smaller value of 1.0 as the updated dynamic weight. The dynamic weight of this relationship edge increases from 0.92 to 1.0, indicating that the knowledge has reached the highest confidence level after user verification. If the current dynamic weight is 0.85, then 0.85 + 0.1 = 0.95, compare 0.95 with 1.0, and take the smaller value of 0.95, updating the dynamic weight to 0.95.
[0063] Step 52: If the feedback information is invalid, reduce the dynamic weight in the weight attribute by the second step value, and compare the reduced dynamic weight with the preset third preset threshold, taking the larger value of the two as the updated dynamic weight.
[0064] As can be seen, step 52 also includes two operations performed sequentially. First, the current value of the dynamic weight is reduced by a second step size. The second step size is a preset positive value used to control the extent of weight reduction caused by each invalid feedback. The second step size can be the same as the first step size, or it can be set to a different value according to actual needs. Then, the reduced dynamic weight is compared with a preset third threshold, and the larger of the two values is taken as the updated dynamic weight. This step is essentially a lower bound truncation mechanism. The third preset threshold is the minimum value that the dynamic weight is allowed to reach, usually set to 0.0, because a weight value below 0 loses its meaning as a confidence level. By taking the larger value, it is ensured that the updated dynamic weight will never fall below the third preset threshold.
[0065] For example, the current dynamic weight of a relation edge is 0.85, the second step size is 0.1, and the third preset threshold is 0.0. When a user reports "invalid," first calculate 0.85 - 0.1 = 0.75, then compare 0.75 with 0.0, and take the larger value of 0.75 as the updated dynamic weight. The dynamic weight of this relation edge decreases from 0.85 to 0.75, indicating that the confidence of this knowledge has decreased after user verification. If after multiple "invalid" feedbacks, the dynamic weight decreases to 0.05, and further reducing it by 0.1 results in -0.05, which is compared with 0.0 and the larger value of 0.0 is taken. The dynamic weight is then updated to 0.0, indicating that this knowledge is marked as completely unreliable.
[0066] Furthermore, the first and second step lengths are key parameters controlling the speed of knowledge evolution. A larger step length results in a greater impact of a single feedback on the weights, leading to faster convergence of knowledge to a stable state, but also greater weight volatility. A smaller step length results in smoother weight changes and higher system stability, but requires more feedback cycles to achieve the same adjustment range. The second and third preset thresholds together define the effective range of dynamic weight values. The first threshold is typically set to 1.0, corresponding to the upper limit of completely reliable confidence; the second threshold is typically set to 0.0, corresponding to the lower limit of completely unreliable confidence. This upper and lower bound setting ensures that the weight values always have mathematical significance as confidence levels and will not exceed a reasonable numerical range due to cumulative updates.
[0067] As can be seen, steps 51 and 52 above provide clear feedback response rules, namely, clear operation paths are set for "valid" and "invalid" feedback respectively, enabling the system to process each user feedback in a deterministic manner and ensuring the predictability and traceability of the knowledge evolution process. The step-by-step incremental update ensures the smoothness of knowledge evolution; that is, each update only changes one step value, avoiding drastic fluctuations in weights due to a single abnormal feedback. This makes the trend of weight changes gradual and predictable, and will not disrupt the stability of the knowledge graph due to individual user misoperations or extreme situations.
[0068] In an optional embodiment of this application, the method of determining the solution to be output from the sorted solutions based on the urgency of the fault information involved in step 103 above may further include: Step 61: If the urgency of the fault information is the highest urgency level, output the sorted solutions within a preset time period; among them, output the solution with the highest comprehensive score within the preset time period. In this embodiment, the preset duration is a pre-defined time limit used to constrain the maximum response time of the system in emergency failure scenarios. In industrial sites handling inverter failures, the preset duration is typically set to a millisecond-level time window. In specific application scenarios, the preset duration can be fifty milliseconds, meaning the system must output recommended results to the operator within fifty milliseconds to ensure the operator has sufficient time to take action before equipment damage. Due to the extremely short preset duration, the system cannot wait for the return of cloud computing results in emergency failure scenarios (cloud round trips can take hundreds of milliseconds or even seconds). Therefore, the system must rely on a local knowledge graph deployed at the edge to independently complete the entire retrieval, matching, scoring, and ranking process, and output the results within the preset duration.
[0069] Furthermore, the output of the ranked solution can be configured to either provide a complete list of ranked solutions under the first level of urgency, or only the top-ranked solution, depending on actual needs. For example, in an emergency, operators often lack the time to compare and select from multiple solutions, typically only the solution with the highest overall score is output to reduce decision-making burden and facilitate immediate execution. While the description of the ranked solution in step 61 is consistent with that in step 103, those skilled in the art will understand that in emergency scenarios, the actual output of the ranked solution can be the optimal solution itself, or a list of optimal solutions and alternatives with complete ranking information.
[0070] In addition, in this emergency failure scenario, although recommendation results were output within the preset time, a summary of the failure was simultaneously uploaded to the cloud asynchronously. Upon receiving the data, the cloud utilizes its global knowledge graph and more sophisticated analytical capabilities to verify the recommendation conclusions from the edge. When the cloud verifies that the edge recommendation is correct, this recommendation knowledge is marked as a high-confidence sample to strengthen the weights of relevant relationship edges; when the cloud verifies that the edge recommendation is inaccurate or that a better solution exists, the cloud records the analysis result for reference during subsequent knowledge graph updates. It should be noted that since the handling of the emergency failure has already been completed under the guidance of the edge recommendation results, the cloud verification results are generally not used to correct the current output, but rather to optimize the recommendation quality for subsequent similar scenarios.
[0071] Step 62: If the urgency level of the fault information is the second urgency level, determine the highest comprehensive score of the solution in the recommended set as the highest comprehensive score. If the highest comprehensive score is greater than or equal to the fourth preset threshold, directly output the sorted solutions; wherein, the urgency level of the first urgency level is higher than the urgency level of the second urgency level. As can be seen, in this embodiment, the urgency level is divided into at least two levels. The first urgency level corresponds to fault types with a high degree of harm, such as grounding faults and bus overvoltage faults. If these faults are not addressed immediately, they may cause serious equipment damage, fire, or personal injury. Therefore, the system needs to provide handling suggestions as quickly as possible, and additional delays due to waiting for cloud computing are not allowed. The second urgency level corresponds to fault types with a lower degree of harm. For example, an "overcurrent fault" may only require parameter adjustment to recover under light load conditions, and a "communication anomaly" may only require checking the communication line. These faults pose a less direct threat to life and property safety, and the response time requirement can be appropriately relaxed while ensuring the accuracy of the recommendations. This allows the system to utilize global knowledge from the cloud for more in-depth analysis when necessary.
[0072] Furthermore, in this embodiment, the fourth preset threshold is a pre-set confidence threshold used to determine whether the reliability of the edge-end recommendation result is sufficient for direct adoption. When the highest comprehensive score reaches or exceeds this threshold, it indicates that the edge-end's diagnosis and recommendation of the fault have a high degree of confidence, and the result can be directly output without waiting for cloud intervention. In a specific application scenario, the fourth preset threshold is set to 0.8. This value is selected based on the following considerations: when the comprehensive score is between 0 and 1, a score above 0.8 means that the recommendation scheme has high reliability and the risk of direct adoption is low.
[0073] Step 63: If the highest comprehensive score is less than the fourth preset threshold, the fault information, real-time operating condition data and matching candidate relationship edges are sent to the cloud, and the solution generated by the cloud after performing conditional retrieval and comprehensive score calculation based on the global knowledge graph is received, and the solution generated by the cloud is used as the output result.
[0074] It should be noted that the pre-built knowledge graph mentioned above is a knowledge graph deployed and running on the local machine of the frequency converter or on a workshop-level edge server, and can be referred to as an "edge knowledge graph" or "local knowledge graph" in this application. This knowledge graph stores the core knowledge necessary for frequency converter fault diagnosis, including fault entities, operating condition parameter entities, environmental parameter entities, solution entities, and quadruple relationship edges connecting these entities. Each relationship edge is accompanied by constraint interval attributes and weight attributes. The size of the edge knowledge graph is typically controlled within a few thousand entity nodes and tens of thousands of relationship edges to ensure millisecond-level retrieval and reasoning on edge computing devices. The aforementioned "global knowledge graph" is a cloud-based knowledge graph with a structure identical to the edge knowledge graph. It also employs a four-tuple structure and includes fault entities, operating condition parameter entities, environmental parameter entities, solution entities, and relational edges with constrained intervals and weighted attributes. The core difference lies in the scope of knowledge coverage: the global knowledge graph is the parent or superset of the edge knowledge graph, encompassing not only all its content but also aggregating massive amounts of fault handling cases and expert experience from different regions, factories, and inverter models. The difference between the edge and global knowledge graphs... The system employs a collaborative mechanism of "regular synchronization + on-demand updates." During system initialization and deployment, the edge knowledge graph retrieves a subset of knowledge related to the local device model and application scenario from the global knowledge graph. During system operation, dynamic weights accumulated at the edge nodes through user feedback are periodically synchronized to the cloud to update the dynamic weights of corresponding relationship edges in the global knowledge graph. Simultaneously, valid feedback and new knowledge from other edge nodes, aggregated in the cloud, are integrated into the global knowledge graph and synchronized to each edge node on demand, enabling continuous enrichment of the global knowledge base and iterative updates to each edge knowledge graph.
[0075] In response, when the highest comprehensive score falls below the fourth preset threshold, it indicates that the confidence level of all recommended solutions obtained by the edge terminal based on the local knowledge graph is insufficient. In this case, directly outputting recommended solutions with insufficient confidence could lead to users adopting incorrect solutions, delaying fault handling, or even causing more serious equipment damage. Therefore, relevant data is sent to the cloud for more comprehensive analysis and calculation using the cloud's global knowledge graph. The data received by the cloud includes the current fault code, real-time operating data, and the set of candidate relationship edges that have been successfully matched by the edge terminal. The cloud re-executes conditional retrieval and comprehensive score calculation based on its global knowledge graph. Because the global knowledge graph contains more relationship edges (such as similar cases from other factories, the latest updated expert knowledge, etc.), the cloud retrieval may obtain a richer set of candidate relationship edges than the edge terminal, thus obtaining recommended solutions with higher comprehensive scores, or providing alternative solutions that were originally missing by the edge terminal. After the cloud calculation is completed, the generated solution is sent back to the edge terminal. The edge terminal presents the solution sent back from the cloud as the output result to the user.
[0076] For example, an inverter reports a "communication anomaly" fault, classified as a second-urgency fault. After the edge terminal performs a search and matching, due to the limited number of candidate edges related to "communication anomaly" in the local knowledge graph (only two, both with low confidence), the top-ranked solution has a comprehensive score of only 0.65, lower than the fourth preset threshold of 0.8. The system determines that the edge terminal's confidence is insufficient and sends the current fault code, real-time operating data, and matched candidate edges to the cloud. The cloud searches its global knowledge graph and finds that there is actually a third solution for communication anomalies under similar operating conditions for this inverter model (from a recently uploaded successful case from another factory). After comprehensive scoring, this new solution has a comprehensive score of 0.91. The cloud sends this solution back to the edge terminal, which uses the solution generated by the cloud as the final output. Simultaneously, the solution returned by the cloud and its associated edge information are stored in the edge terminal's local cache for subsequent retrieval and matching of similar faults.
[0077] As can be seen, steps 61 to 63 above achieve a balance between response speed and recommendation accuracy. For emergency faults, millisecond-level response is given the highest priority, ensuring that operators receive handling suggestions before equipment damage occurs. For routine faults, recommendation accuracy is given the highest priority, and cloud-based global knowledge is used to enhance the recommendation when edge-side confidence is insufficient. Furthermore, efficient use of cloud resources is achieved by setting a fourth preset threshold, requesting cloud intervention only when edge-side confidence is insufficient. In most routine fault scenarios, the edge-side knowledge graph is sufficient to provide high-confidence recommendations, without consuming cloud computing resources and network bandwidth; cloud-based global knowledge is only used to supplement edge-side knowledge when it is insufficient. This design significantly improves the overall resource utilization efficiency of the system.
[0078] The present application will be explained and described below with reference to specific embodiments, which provide a method and system for constructing a four-tuple conditional knowledge graph for inverter fault diagnosis, such as... Figure 2 The system shown includes a data acquisition layer, a knowledge construction and feature processing layer, a recommendation engine layer, a feedback evolution layer, an application display layer, and an edge computing module.
[0079] The data acquisition layer is used to collect multi-source data from the frequency converter through sensor networks, data interfaces, and manual input: sensor data (vibration, current, temperature), log work order data, fault waveform files, operating parameters (load rate, speed, running time, bus voltage), and environmental data (temperature, humidity, altitude). Each acquisition unit is connected to the data bus of the data acquisition layer via an industrial Ethernet network, and the sampling frequency is dynamically adjusted according to the signal type.
[0080] The knowledge construction and feature processing layer is used to transform the collected multi-source data into structured knowledge, including the four-element knowledge graph construction unit, entity extraction and relation mining unit.
[0081] The recommendation engine layer is used for dynamic reasoning and recommendation based on real-time conditions, including conditional retrieval units, fusion ranking and confidence scoring units.
[0082] The feedback evolution layer is used to collect user feedback and update the weights of relation edges in the knowledge graph.
[0083] The application presentation layer provides web interfaces, mobile applications, and API services to display recommendation results and receive user feedback.
[0084] Edge computing modules are deployed locally on the inverter or on workshop-level edge servers for rapid response to lightweight inference tasks.
[0085] Furthermore, such as Figure 3 The illustrated quadruple-based conditional knowledge graph structure includes entities for faults, operating parameters, environmental factors, and solutions. Each relation edge is accompanied by operating condition constraint interval attributes and environmental constraint interval attributes. During retrieval, based on real-time collected operating and environmental parameters, only solutions pointed to by relation edges whose constraint intervals contain the current actual values are returned. A solution recommendation interface is provided, allowing users to rate solutions based on their actual effectiveness. The system dynamically adjusts the weights of the corresponding quadruple edges based on user ratings.
[0086] The entity definition includes: (1) Fault code entities include fault types such as overcurrent OC, overvoltage OV, overheat OH, grounding GF, and communication abnormality CE. Each fault code entity is assigned a fault severity level (high / medium / low) attribute. (2) The operating condition parameters include load rate, speed, running time, radiator temperature, bus voltage, etc., and each parameter is attached with a normal range threshold and sampling frequency attribute; (3) The solution entity includes parameter adjustment steps, component replacement process, software upgrade method, hardware reset operation, etc., with additional operation complexity (low / medium / high) and expected time consumption attributes; (4) Environmental factors include ambient temperature, humidity, altitude, and electromagnetic interference level.
[0087] The relationship edges are defined using a four-tuple conditional matching relationship. Taking overcurrent fault as an example: when the load rate is below 30% and the ambient temperature is below 40℃, it is recommended to "increase the acceleration time". When the load rate is above 80% and the ambient temperature is above 40℃, it is recommended to "replace the IGBT module". These two matching relationships are stored as two independent four-tuple edges in the graph, and the attribute fields of each edge are shown in Table 1. Table 1
[0088] Furthermore, such as Figure 4 The conditional search process based on real-time operating conditions shown includes the following steps: Step 401: When the frequency converter reports fault code F_cur, the system synchronously obtains the current operating condition parameter set C_cur (e.g., load rate = 25%) and environmental parameter E_cur (e.g., ambient temperature = 28℃) from the data acquisition layer.
[0089] Step 402: Using F_cur as the key, extract all quadruple relation edges from the knowledge graph that originate from the faulty entity.
[0090] Step 403: Traverse each relation edge and check whether C_cur falls within the working condition constraint range of that edge and whether E_cur falls within the environmental constraint range.
[0091] Step 404: Add the solution entity pointed to by the relation edge that satisfies both interval conditions to the candidate set, and record the static weight and dynamic weight of the edge.
[0092] Step 405: Is the candidate set empty? If yes, proceed to step 406; otherwise, proceed to step 407. The solutions in the candidate set are sorted and output according to the comprehensive score = 0.4 × static weight + 0.6 × dynamic weight.
[0093] Step 406, perform relaxation matching: calculate the distance between the midpoint of the constraint interval of each edge and the current actual value, and select the top 3 edges with the smallest distances for output.
[0094] For example, when the current load rate is 25% and the temperature is 28℃, edge 1 satisfies the condition (25∈[0,30], 28∈[-10,40]), while edge 2 does not satisfy (25...). [80,100]). The system output increases the acceleration time, with a confidence level of 0.85.
[0095] When the current load rate is 90% and the temperature is 45℃, edge 1 does not meet the requirements, while edge 2 does. The system outputs "Replace IGBT module" with a confidence level of 0.92.
[0096] Step 407: The candidate solutions are sorted and output according to their comprehensive scores.
[0097] After each recommendation is completed, user feedback is collected on the interface, including two options: valid or invalid.
[0098] If the user marks it as valid, the system locates the quadruple edge that the recommendation hits and executes: New dynamic weight = min(1.0, Old dynamic weight + 0.1). If the user marks it as "invalid," it executes: New dynamic weight = max(0.0, Old dynamic weight - 0.1). The updated dynamic weight takes effect immediately in subsequent queries. Since the dynamic weight is bound to a specific quadruple edge, the update only affects the recommendation ranking under that specific working condition. For example, if a user submits invalid feedback for increasing acceleration time under light load conditions, the weight of this solution will only decrease under light load conditions; the weight of this solution will remain unaffected under heavy load conditions.
[0099] like Figure 5 As shown, the system adopts an edge-cloud hierarchical collaborative architecture: The edge layer is deployed locally on the inverter or on workshop-level edge servers, storing a complete four-tuple knowledge graph (typically ranging from a few thousand nodes to tens of thousands of edges) and performing conditional retrieval. The cloud layer deploys a complete backup of the global knowledge graph, as well as deep analysis capabilities for complex faults.
[0100] The reasoning process is as follows: After a fault event is triggered, the edge layer first determines whether it is an emergency fault based on the preset severity level of the fault code.
[0101] If it is an emergency fault (such as "grounding" or "bus overvoltage"), the edge layer immediately performs a conditional search based on the local knowledge graph and outputs a preliminary recommendation result (such as "emergency power disconnection") within 50ms. At the same time, the fault summary is asynchronously uploaded to the cloud for verification.
[0102] In the case of a routine fault, the edge layer performs conditional retrieval and calculates the confidence score (i.e., the overall score). If the confidence score is greater than or equal to 0.8, the recommendation result is output directly; if the confidence score is less than 0.8 or the relevant information is missing in the local cache, the edge module actively requests the cloud to perform complete inference. After the cloud completes the inference, it sends the result back to the edge layer and updates the edge's local cache.
[0103] like Figure 6 The recommendation results interface shown below displays the following content: Recommended Solution List: Multiple recommended solutions are displayed, sorted by overall score, with each solution accompanied by a confidence score.
[0104] Matching criteria explanation: This shows which constraint range the current operating condition value falls within. For example, the current load rate is 25%, which is within the applicable range [0%, 30%] of recommended scheme A; the current temperature is 28℃, which is within the applicable range [-10℃, 40℃].
[0105] Confidence score composition: This section shows the calculation process of the overall score, including static weights, dynamic weights, and their proportions.
[0106] Feedback Collection Area: Includes an "Valid / Invalid" button and a performance rating slider for collecting user feedback.
[0107] As can be seen, this application adopts a modeling approach using quadruples (fault entity, operating condition entity, environment entity, and solution entity) plus explicit constraint attributes, attaching applicable operating condition and environment intervals to each matching relationship. Compared to existing triplet models, this application can fully preserve the multivariate constraints in fault diagnosis, avoiding information distortion. In typical inverter fault scenarios, the accuracy of conditional retrieval can be significantly improved, while the retrieval calculation only involves numerical interval judgment, and the edge inference latency can be controlled within 50ms. In addition, each quadruple edge is an independent evaluable unit, and user feedback can accurately locate a specific edge under specific operating conditions, achieving fine-grained, loosely coupled local updates without the need to reconstruct the entire graph. Moreover, an edge-cloud collaborative inference strategy is set up with fault severity classification, with emergency faults responding quickly from the edge, and routine faults determining whether to request the cloud based on confidence level. Furthermore, when outputting solutions, the system can intuitively display which relational edge constraint intervals the current operating condition parameters satisfy, allowing users to clearly understand the basis for the recommendations.
[0108] Corresponding to the above Figure 1 This application also provides a device for providing a fault solution for frequency converters based on a knowledge graph, such as... Figure 7 As shown, the device includes: The data acquisition module 702 is used to collect real-time operating data of the frequency converter in response to the fault information reported by the frequency converter. The first processing module 704 is used to filter candidate relation edges associated with fault information from a pre-built knowledge graph based on fault information, and match the constraint interval attributes of each candidate relation edge with real-time operating condition data, so as to add the solution entity corresponding to the successfully matched candidate relation edge to the recommendation set; wherein, the knowledge graph includes multiple types of entities and multiple relation edges connecting entities, each relation edge is associated with a fault entity, an operating condition parameter entity, an environmental parameter entity and a solution entity, and each relation edge is attached with a constraint interval attribute and a weight attribute; The second processing module 706 is used to determine the comprehensive score of each solution in the recommendation set according to the weight attribute of each candidate relation edge, sort the solutions in the recommendation set from high to low according to the comprehensive score, and determine the solution to be output from the sorted solutions according to the urgency of the fault information.
[0109] In an optional embodiment of this application, the first processing module in this application includes: The matching unit is used to match the operating parameters in the real-time operating data with the operating constraint interval in the constraint interval attribute, and to match the environmental parameters in the real-time operating data with the environmental constraint interval in the constraint interval attribute. The first processing unit is used to add the solution entities corresponding to the successfully matched candidate relation edges to the recommendation set.
[0110] In an optional embodiment of this application, the first processing unit in this application includes: The first determining sub-unit is used to determine that the real-time operating data and the candidate relationship edge are successfully matched when the operating parameters fall within the operating condition constraint range and the environmental parameters fall within the environmental constraint range. The second determining subunit is used to determine the first difference between the operating condition parameter and the midpoint of the operating condition constraint interval of each candidate relation edge, and to determine the second difference between the environmental parameter and the midpoint of the environmental constraint interval of each candidate relation edge, when the operating condition parameter does not fall within the operating condition constraint interval or the environmental parameter does not fall within the environmental constraint interval. The third determining subunit determines the comprehensive deviation degree between each candidate relation edge and the real-time operating condition data based on the first difference and the second difference, and determines that the candidate relation edge with a comprehensive deviation degree less than the first preset threshold is successfully matched with the real-time operating condition data; Add a sub-unit to add the solution entity corresponding to the successfully matched candidate relation edge to the recommendation set.
[0111] In an optional embodiment of this application, the second processing module includes: The second processing unit is used to multiply the preset first weighting coefficient with the static weight in the weight attribute to obtain the first product result; The third processing unit is used to multiply the preset second weighting coefficient with the dynamic weight in the weight attribute to obtain the second product result; The fourth processing unit is used to sum the first product result and the second product result as a comprehensive score.
[0112] In optional embodiments of this application, the apparatus further includes: The receiving module is used to receive feedback information from the user regarding the sorted solutions in the current output; The update module is used to perform incremental updates on the weight attributes corresponding to the successfully matched candidate relation edges based on the feedback information.
[0113] In an optional embodiment of this application, the update module includes: The first update unit is used to increase the dynamic weight in the weight attribute by a first step length value if the feedback information is valid, and compare the increased dynamic weight with a preset second preset threshold, and take the smaller value of the two as the updated dynamic weight. The second update unit is used to reduce the dynamic weight in the weight attribute by a second step size if the feedback information is invalid, and compare the reduced dynamic weight with a preset third preset threshold, taking the larger value of the two as the updated dynamic weight.
[0114] In an optional embodiment of this application, the second processing module includes: The first output unit is used to output sorted solutions within a preset time period when the urgency of the fault information is the highest; among which, the solution with the highest comprehensive score is output within the preset time period. The second output unit is used to determine the highest comprehensive score of the solution in the recommended set as the highest comprehensive score when the urgency level of the fault information is the second urgency level. If the highest comprehensive score is greater than or equal to a fourth preset threshold, the sorted solutions are directly output. The urgency level of the first urgency level is higher than that of the second urgency level. The fifth processing unit is used to send fault information, real-time operating condition data and matching candidate relationship edges to the cloud if the highest comprehensive score is less than the fourth preset threshold, and to receive the solution generated by the cloud after performing conditional retrieval and comprehensive score calculation based on the global knowledge graph, and to use the solution generated by the cloud as the output result.
[0115] like Figure 8As shown in the figure, this application provides an electronic device including a processor 811, a communication interface 812, a memory 813, and a communication bus 814, wherein the processor 811, the communication interface 812, and the memory 813 communicate with each other through the communication bus 814. Memory 813 is used to store computer programs; In one embodiment of this application, when the processor 811 executes the program stored in the memory 813, it implements the method of providing a knowledge graph-based output inverter fault solution provided in any of the aforementioned method embodiments. Its function is similar and will not be described again here.
[0116] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method for providing a knowledge graph-based output inverter fault solution as provided in any of the foregoing method embodiments.
[0117] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.
[0118] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0119] It should be understood that the terminology used herein is for the purpose of describing particular exemplary embodiments only and is not intended to be limiting. Unless the context clearly indicates otherwise, the singular forms “a,” “an,” and “described” as used herein may also include the plural forms. The terms “comprising,” “including,” “containing,” and “having” are inclusive and therefore indicate the presence of the stated features, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, elements, components, and / or combinations thereof. The method steps, processes, and operations described herein are not construed as requiring them to be performed in a particular order described or illustrated unless the order of performance is explicitly indicated. It should also be understood that additional or alternative steps may be used.
[0120] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.
Claims
1. A method of outputting a variable frequency drive fault resolution solution based on a knowledge graph, the method comprising: include: In response to the fault information reported by the frequency converter, the real-time operating condition data of the frequency converter is collected; Based on the fault information, candidate relationship edges associated with the fault information are filtered from a pre-constructed knowledge graph. The real-time operating condition data is then matched with the constraint interval attributes of each candidate relationship edge to add the solution entity corresponding to the successfully matched candidate relationship edge to the recommendation set. The knowledge graph includes multiple types of entities and multiple relationship edges connecting the entities. Each relationship edge is associated with a fault entity, an operating condition parameter entity, an environmental parameter entity, and a solution entity, and each relationship edge is attached with constraint interval attributes and weight attributes. The comprehensive score of each solution in the recommendation set is determined based on the weight attribute of each candidate relation edge, and the solutions in the recommendation set are sorted from high to low according to the comprehensive score. The solution to be output is determined from the sorted solutions according to the urgency of the fault information.
2. The method of claim 1, wherein, The real-time operating data is matched with the constraint interval attributes of each candidate relationship edge to add the solution entity corresponding to the successfully matched candidate relationship edge to the recommendation set, including: Match the operating parameters in the real-time operating data with the operating constraint interval in the constraint interval attribute, and match the environmental parameters in the real-time operating data with the environmental constraint interval in the constraint interval attribute. Add the solution entity corresponding to the successfully matched candidate relation edge to the recommendation set.
3. The method of claim 2, wherein, Add the solution entity corresponding to the successfully matched candidate relation edge to the recommendation set, including: If the operating condition parameters fall within the operating condition constraint range and the environmental parameters fall within the environmental constraint range, it is determined that the real-time operating condition data and the candidate relationship edge are successfully matched. If the operating condition parameter does not fall within the operating condition constraint interval, or the environmental parameter does not fall within the environmental constraint interval, determine the first difference between the operating condition parameter and the midpoint of the operating condition constraint interval of each candidate relation edge, and determine the second difference between the environmental parameter and the midpoint of the environmental constraint interval of each candidate relation edge. Based on the first difference and the second difference, the comprehensive deviation degree between each candidate relationship edge and the real-time operating condition data is determined, and the candidate relationship edge whose comprehensive deviation degree is less than the first preset threshold is successfully matched with the real-time operating condition data. Add the solution entity corresponding to the successfully matched candidate relation edge to the recommendation set.
4. The method of claim 1, wherein, The comprehensive score of each solution in the recommendation set is determined based on the weight attributes of each candidate relation edge, including: The first product result is obtained by multiplying the preset first weighting coefficient with the static weight in the weight attribute; The second product result is obtained by multiplying the preset second weighting coefficient with the dynamic weight in the weight attribute; The sum of the first product result and the second product result is used as the comprehensive score.
5. The method of claim 1, wherein, The method further includes: Receive feedback from users regarding the sorted solutions in the current output; Based on the feedback information, the weight attributes corresponding to the successfully matched candidate relation edges are incrementally updated.
6. The method of claim 5, wherein, Based on the feedback information, incremental updates are performed on the weight attributes corresponding to the successfully matched candidate relation edges, including: If the feedback information is valid, the dynamic weight in the weight attribute is increased by the first step length value, and the increased dynamic weight is compared with the preset second preset threshold. The smaller of the two values is taken as the updated dynamic weight. If the feedback information is invalid, the dynamic weight in the weight attribute is reduced by a second step value, and the reduced dynamic weight is compared with a preset third preset threshold. The larger of the two values is taken as the updated dynamic weight.
7. The method of claim 1, wherein, Based on the urgency of the fault information, the solution to be output is determined from the sorted solutions, including: If the urgency level of the fault information is the highest, the sorted solutions will be output within a preset time period; wherein, the solution with the highest comprehensive score will be output within the preset time period. When the urgency of the fault information is the second urgency level, the highest comprehensive score of the solution in the recommendation set is determined as the highest comprehensive score. If the highest comprehensive score is greater than or equal to a fourth preset threshold, the sorted solutions are directly output. The urgency of the first urgency level is higher than that of the second urgency level. If the highest comprehensive score is less than the fourth preset threshold, the fault information, the real-time operating condition data and the matched candidate relationship edges are sent to the cloud, and the solution generated by the cloud after performing conditional retrieval and comprehensive score calculation based on the global knowledge graph is received, and the solution generated by the cloud is used as the output result.
8. An apparatus for outputting a variable frequency drive fault resolution solution based on a knowledge graph, the apparatus comprising: include: The data acquisition module is used to collect real-time operating data of the frequency converter in response to the fault information reported by the frequency converter. The first processing module is used to filter candidate relationship edges associated with the fault information from a pre-constructed knowledge graph based on the fault information, and match the real-time operating condition data with the constraint interval attributes of each candidate relationship edge, so as to add the solution entity corresponding to the successfully matched candidate relationship edge to the recommendation set; wherein, the knowledge graph includes multiple types of entities and multiple relationship edges connecting the entities, each relationship edge is associated with a fault entity, an operating condition parameter entity, an environmental parameter entity and a solution entity, and each relationship edge is attached with a constraint interval attribute and a weight attribute; The second processing module is used to determine the comprehensive score of each solution in the recommendation set according to the weight attribute of each candidate relation edge, sort the solutions in the recommendation set from high to low according to the comprehensive score, and determine the solution to be output from the sorted solutions according to the urgency of the fault information.
9. An electronic device, comprising: include: The processor, communication interface, memory, and communication bus are connected, with the processor, communication interface, and memory communicating with each other via the communication bus. The memory is used to store computer programs; the processor is used to execute the computer programs to implement the method for outputting inverter fault solutions based on knowledge graphs as described in any one of claims 1-7.
10. A storage medium having stored thereon a computer program, characterized in that When executed by a processor, the computer program implements the method for outputting inverter fault solutions based on knowledge graphs as described in any one of claims 1-7.
