Fault root cause prediction method and device, model training method and device, electronic equipment, storage medium and computer program product

Abstract

The invention relates to a fault root cause prediction method and device, a model training method and device, electronic equipment, a storage medium and a computer program product, and the method comprises the steps: obtaining a fault description and a super-relation knowledge graph, and enabling the fault description to comprise at least one piece of fault phenomenon information, the super-relation knowledge graph is pre-constructed based on a fault case and a preset knowledge graph, the fault case comprises at least one piece of fault phenomenon information, the preset knowledge graph comprises a non-root cause entity and a root cause entity, and the super-relation knowledge graph comprises a case hyperedge and a root cause class hyperedge; the case hyperedge comprises the non-root cause entity or the root cause entity corresponding to each piece of fault phenomenon information in the fault case, and the root cause class hyperedge comprises the root cause entities of the same class obtained by clustering; and inputting the fault description and the super-relation knowledge graph into a fault root cause prediction model to obtain a predicted value of the fault root cause.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence, and specifically to a method for predicting the root causes of failures, a model training method, an apparatus, an electronic device, a storage medium, and a computer program product. Background Technology

[0002] The safe and reliable operation of manufacturing processes is crucial for the sustainable development and long-term profitability of modern industrial production. Production processes involve thousands of process control variables and product quality indicators. Abnormal changes in certain measured variables can trigger a chain reaction in other related variables, leading to quality and performance anomalies. Root cause analysis aims to analyze the causal relationships between failure variables and locate the root cause—that is, to trace the root cause of a failure based on multifaceted failure information. It plays a pivotal role in modern industrial production processes, such as semiconductor manufacturing and repair, automobile manufacturing and repair, and automated control systems for blast furnaces and rolling mills in metallurgy, demonstrating significant commercial value.

[0003] If we follow the approach of knowledge graphs and deep neural networks to diagnose root causes of faults, the prediction of root causes of faults mainly relies on semantic similarity to link entities. Therefore, the importance of semantic feature representation is particularly important. However, the methods in related technologies only extract semantic features based on pre-trained models and do not further optimize feature representation. Summary of the Invention

[0004] The fault root cause prediction method, model training method, apparatus, electronic device, storage medium, and computer program product provided by the exemplary embodiments of this disclosure can at least solve the above-mentioned technical problems and other technical problems not mentioned above.

[0005] According to one aspect of this disclosure, a method for predicting the root cause of a fault is provided. The method includes: acquiring a fault description and a hyperrelationship knowledge graph, wherein the fault description includes at least one fault phenomenon information, the hyperrelationship knowledge graph is pre-constructed based on fault cases and a preset knowledge graph, the fault cases include at least one fault phenomenon information, the preset knowledge graph includes non-root cause entities and root cause entities, the hyperrelationship knowledge graph includes case hyperedges and root cause class hyperedges, the case hyperedges include the non-root cause entity or the root cause entity corresponding to each fault phenomenon information in the fault case, and the root cause class hyperedges include root cause entities of the same class obtained by clustering; and inputting the fault description and the hyperrelationship knowledge graph into a fault root cause prediction model to obtain predicted values ​​for the fault root cause.

[0006] Optionally, the root cause prediction model includes multiple learnable parameter matrices. The step of inputting the fault description and the hyperrelationship knowledge graph into the root cause prediction model to obtain the predicted value of the root cause includes: performing the following operations through the root cause prediction model: extracting global case features of the case hyperedge and global root cause class features of the root cause class hyperedge using a first and a second learnable parameter matrix, respectively; obtaining the predicted value of the root cause class based on the dot product of the global case features and the global root cause class features; obtaining local case features by filtering the global case features using a third learnable parameter matrix; obtaining local root cause features by filtering the global root cause class features using a fourth learnable parameter matrix; adjusting the weights of the corresponding root causes in the local root cause features based on the predicted value of the root cause class to obtain weight-adjusted local root cause features; and obtaining the predicted value of the root cause based on the dot product between the local case features and the weight-adjusted local root cause features.

[0007] Optionally, the filtering of the global root cause class features based on the fourth learnable parameter matrix to obtain local root cause features includes: expanding the global root cause class features corresponding to the predicted values ​​of the root cause class based on the number of root causes in the super-relationship knowledge graph to obtain first global root cause class features; obtaining first structural features of root cause entities in the predicted values ​​of the root cause class based on the index of root causes in the super-relationship knowledge graph; and concatenating and filtering the first global root cause class features and the first structural features based on the fourth learnable parameter matrix to obtain the local root cause features.

[0008] Optionally, the hyperrelationship knowledge graph is constructed through the following steps: determining candidate entities corresponding to the fault phenomenon information from the preset knowledge graph, wherein the candidate entities are a set of entities distinguished according to ontology classes; determining the non-root cause entity or the root cause entity corresponding to the fault phenomenon information based on the similarity between the text features of the fault phenomenon information and the text features of each entity in the candidate entities corresponding to the fault phenomenon information, so as to obtain the case hyperedge of the hyperrelationship knowledge graph; clustering based on the text features of the root cause entities in the preset knowledge graph, and taking the root cause entities of the same cluster as the entities included in the root cause class hyperedge of the hyperrelationship knowledge graph.

[0009] Optionally, the step of extracting the global case features of the case hyperedge and the global root cause features of the root cause hyperedge using the first and second learnable parameter matrices respectively includes: extracting the text features of the fault case and the structural features within the case hyperedge, and obtaining the global case features of the case hyperedge based on the extracted text features and structural features combined with the first learnable parameter matrix; and obtaining the global root cause features of the root cause hyperedge by extracting the text features and structural features within the root cause hyperedge, and obtaining the global root cause features of the root cause hyperedge based on the extracted text features and structural features combined with the second learnable parameter matrix.

[0010] According to another aspect of this disclosure, a training method for a fault root cause prediction model is also provided. The training method includes: acquiring training data and a hyperrelationship knowledge graph, wherein the training data includes fault cases for training, each fault case containing at least one fault phenomenon information; the hyperrelationship knowledge graph is pre-constructed based on the fault cases and a preset knowledge graph, the preset knowledge graph containing non-root cause entities and root cause entities; the hyperrelationship knowledge graph containing case hyperedges and root cause class hyperedges, each case hyperedge containing the non-root cause entity or the root cause entity corresponding to each fault phenomenon information in the fault case; and the root cause class hyperedges containing root cause entities of the same class obtained through clustering; inputting the fault cases and the hyperrelationship knowledge graph into the fault root cause prediction model to obtain predicted values ​​of the fault root causes; calculating a first loss between the predicted values ​​of the fault root causes and their corresponding true values; and training the fault root cause prediction model based on the first loss.

[0011] Optionally, the root cause prediction model includes multiple learnable parameter matrices, wherein inputting the fault case and the hyperrelationship knowledge graph into the root cause prediction model to obtain the predicted value of the root cause includes: performing the following operations through the root cause prediction model: extracting global case features of the case hyperedge and global root cause class features of the root cause class hyperedge using the first and second learnable parameter matrices, respectively; obtaining the predicted value of the root cause class of the fault based on the dot product of the global case features and the global root cause class features; and performing the prediction on the global case features based on the third learnable parameter matrix. The process involves filtering the features to obtain local case features; filtering the global root cause class features based on the fourth learnable parameter matrix to obtain local root cause features; adjusting the weights of the corresponding root causes in the local root cause features based on the predicted values ​​of the root cause classes to obtain weight-adjusted local root cause features; and obtaining the predicted value of the fault root cause based on the dot product between the local case features and the weight-adjusted local root cause features. The step of training the fault root cause prediction model based on the first loss includes: training the plurality of learnable parameter matrices in the fault root cause prediction model based on the first loss.

[0012] Optionally, the method further includes: calculating a second loss between the predicted value and the corresponding true value of the root cause class; wherein, training the fault root cause prediction model based on the first loss includes: training the plurality of learnable parameter matrices in the fault root cause prediction model based on the first loss and the second loss.

[0013] Optionally, the method further includes: calculating a third loss through contrastive learning based on the similarity between the text features of the fault case and the text features of the root cause entity in the hyperrelationship knowledge graph; training the fault root cause prediction model based on the first loss includes: training the plurality of learnable parameter matrices in the fault root cause prediction model based on the first loss and the third loss.

[0014] Optionally, the method further includes: calculating a third loss through contrastive learning based on the similarity between the text features of the fault case and the text features of the root cause entity in the hyperrelationship knowledge graph; wherein, training the plurality of learnable parameter matrices in the fault root cause prediction model based on the first loss and the second loss includes: training the plurality of learnable parameter matrices in the fault root cause prediction model based on the first loss, the second loss and the third loss.

[0015] According to another aspect of the embodiments of this disclosure, a fault root cause prediction device is also provided. The fault root cause prediction device includes: a data acquisition unit configured to: acquire a fault description and a super-relationship knowledge graph, wherein the fault description includes at least one fault phenomenon information, the super-relationship knowledge graph is pre-constructed based on fault cases and a preset knowledge graph, the fault case includes at least one fault phenomenon information, the preset knowledge graph includes non-root cause entities and root cause entities, the super-relationship knowledge graph includes case superedges and root cause class superedges, the case superedges include the non-root cause entity or the root cause entity corresponding to each fault phenomenon information in the fault case, and the root cause class superedges include root cause entities of the same class obtained by clustering; and a root cause prediction unit configured to: input the fault description and the super-relationship knowledge graph into a fault root cause prediction model to obtain a predicted value of the fault root cause.

[0016] Optionally, the root cause prediction model includes multiple learnable parameter matrices, wherein the root cause prediction unit is configured to perform the following operations through the root cause prediction model: extracting global case features of the case hyperedge and global root cause class features of the root cause class hyperedge using a first learnable parameter matrix and a second learnable parameter matrix, respectively; obtaining the predicted value of the root cause class of the fault root cause based on the dot product of the global case features and the global root cause class features; obtaining local case features by filtering the global case features based on a third learnable parameter matrix; obtaining local root cause features by filtering the global root cause class features based on a fourth learnable parameter matrix; adjusting the weights of the corresponding root causes in the local root cause features based on the predicted value of the root cause class to obtain weight-adjusted local root cause features; and obtaining the predicted value of the fault root cause based on the dot product between the local case features and the weight-adjusted local root cause features.

[0017] Optionally, the root cause prediction unit is configured to: expand the global root cause class features corresponding to the predicted value of the root cause class based on the number of root causes in the super-relationship knowledge graph to obtain a first global root cause class feature; obtain a first structural feature of the root cause entity in the predicted value of the root cause class based on the index of the root cause in the super-relationship knowledge graph; and concatenate and filter the first global root cause class feature and the first structural feature based on the fourth learnable parameter matrix to obtain the local root cause feature.

[0018] Optionally, the hyperrelationship knowledge graph is constructed through the following steps: determining candidate entities corresponding to the fault phenomenon information from the preset knowledge graph, wherein the candidate entities are a set of entities distinguished according to ontology classes; determining the non-root cause entity or the root cause entity corresponding to the fault phenomenon information based on the similarity between the text features of the fault phenomenon information and the text features of each entity in the candidate entities corresponding to the fault phenomenon information, so as to obtain the case hyperedge of the hyperrelationship knowledge graph; clustering based on the text features of the root cause entities in the preset knowledge graph, and taking the root cause entities of the same cluster as the entities included in the root cause class hyperedge of the hyperrelationship knowledge graph.

[0019] Optionally, the root cause prediction unit is configured to: extract text features of the failure case and structural features within the case hyperedge, and obtain global case features of the case hyperedge based on the extracted text features and structural features combined with the first learnable parameter matrix; and obtain global root cause class features of the root cause class hyperedge by extracting text features and structural features within the root cause class hyperedge, and obtaining global root cause class features of the root cause class hyperedge based on the extracted text features and structural features combined with the second learnable parameter matrix.

[0020] According to another aspect of the embodiments of this disclosure, a training apparatus for a fault root cause prediction model is also provided. The training apparatus includes: a data acquisition unit configured to acquire training data and a hyperrelationship knowledge graph, wherein the training data includes fault cases for training, each fault case containing at least one fault phenomenon information, the hyperrelationship knowledge graph being pre-constructed based on the fault cases and a preset knowledge graph, the preset knowledge graph containing non-root cause entities and root cause entities, the hyperrelationship knowledge graph containing case hyperedges and root cause class hyperedges, each case hyperedge containing the non-root cause entity or the root cause entity corresponding to each fault phenomenon information in the fault case, and the root cause class hyperedge containing root cause entities of the same class obtained by clustering; a root cause prediction unit configured to input the fault cases and the hyperrelationship knowledge graph into the fault root cause prediction model to obtain a predicted value of the fault root cause; a first loss calculation unit configured to calculate a first loss between the predicted value of the fault root cause and the corresponding true value; and a model training unit configured to train the fault root cause prediction model based on the first loss.

[0021] Optionally, the root cause prediction model includes multiple learnable parameter matrices, wherein the root cause prediction unit is configured to perform the following operations through the root cause prediction model: extracting global case features of the case hyperedge and global root cause class features of the root cause class hyperedge using a first learnable parameter matrix and a second learnable parameter matrix, respectively; obtaining the predicted value of the root cause class of the fault root cause based on the dot product of the global case features and the global root cause class features; obtaining local case features by filtering the global case features based on a third learnable parameter matrix; obtaining local root cause features by filtering the global root cause class features based on a fourth learnable parameter matrix; adjusting the weights of the corresponding root causes in the local root cause features based on the predicted value of the root cause class to obtain weight-adjusted local root cause features; and obtaining the predicted value of the fault root cause based on the dot product between the local case features and the weight-adjusted local root cause features; wherein the model training unit is configured to train the multiple learnable parameter matrices in the fault root cause prediction model based on the first loss.

[0022] Optionally, the apparatus further includes a second loss calculation unit configured to: calculate a second loss between the predicted value and the corresponding true value of the root cause class; wherein the model training unit is configured to: train the plurality of learnable parameter matrices in the fault root cause prediction model based on the first loss and the second loss.

[0023] Optionally, the apparatus further includes a third loss calculation unit, configured to: calculate a third loss through contrastive learning based on the similarity between the text features of the fault case and the text features of the root cause entity in the hyperrelationship knowledge graph; the model training unit is configured to: train the plurality of learnable parameter matrices in the fault root cause prediction model based on the first loss and the third loss.

[0024] Optionally, the apparatus further includes a third loss calculation unit, configured to: calculate a third loss through contrastive learning based on the similarity between the text features of the fault case and the text features of the root cause entity in the hyperrelationship knowledge graph; wherein the model training unit is configured to: train the plurality of learnable parameter matrices in the fault root cause prediction model based on the first loss, the second loss and the third loss.

[0025] According to another aspect of the present disclosure, an electronic device is also provided, comprising: at least one processor; at least one memory storing computer-executable instructions, wherein, when executed by the at least one processor, the computer-executable instructions cause the at least one processor to execute the training method of the fault root cause prediction model and the fault root cause prediction method as described above.

[0026] According to another aspect of the embodiments of this disclosure, a computer-readable storage medium storing instructions is also provided, which, when executed by at least one processor, cause the at least one processor to perform the training method for the fault root cause prediction model and the fault root cause prediction method as described above.

[0027] According to another aspect of the embodiments of this disclosure, a system is also provided that includes at least one computing device and at least one storage device for storing instructions, wherein, when the instructions are executed by the at least one computing device, they cause the at least one computing device to perform the training method for the root cause prediction model and the prediction method for the root cause of the fault as described above.

[0028] According to another aspect of the embodiments of this disclosure, a computer program product is also provided, including a computer program / instructions, which, when executed by a processor, implement the training method for the fault root cause prediction model and the fault root cause prediction method as described in any of the above.

[0029] The technical solutions provided in this disclosure offer at least the following beneficial effects:

[0030] Based on the fault root cause prediction method, model training method, device, electronic equipment, storage medium and computer program product disclosed herein, fault root cause analysis can be performed by constructing a hyper-relationship knowledge graph, thereby maximizing the utilization of knowledge and experience information, realizing comprehensive consideration of multiple data features, truly conforming to actual production conditions and improving the efficiency of root cause prediction.

[0031] Furthermore, combining coarse-grained and fine-grained root cause prediction can improve the interpretability of knowledge-based deep learning models, enabling their application in actual production.

[0032] In addition, comparative learning can optimize the semantic feature representation of fault information, thereby helping to improve the effectiveness of the main task of fault root cause analysis. Attached Figure Description

[0033] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure, and are not intended to unduly limit this disclosure.

[0034] Figure 1 A flowchart illustrating a training method for a fault root cause prediction model in an exemplary embodiment of this disclosure is provided.

[0035] Figure 2 This illustration shows a flowchart of entity linking in an exemplary embodiment of the present disclosure;

[0036] Figure 3This illustration shows a flowchart of text clustering for fault root causes in an exemplary embodiment of the present disclosure;

[0037] Figure 4 This illustration shows a flowchart of constructing a hyperrelationship knowledge graph in an exemplary embodiment of this disclosure;

[0038] Figure 5 This illustration shows a flowchart of fine-grained root cause prediction based on a hyperrelationship knowledge graph in an exemplary embodiment of this disclosure.

[0039] Figure 6 This illustration shows a flowchart of coarse-grained root cause prediction based on a hyperrelationship knowledge graph in an exemplary embodiment of this disclosure.

[0040] Figure 7 A flowchart illustrating a training method for a fault root cause prediction model in an exemplary embodiment of this disclosure is shown.

[0041] Figure 8 A flowchart illustrating a self-supervised fault root cause analysis framework based on hyperrelationship knowledge graphs in an exemplary embodiment of this disclosure is shown.

[0042] Figure 9 A flowchart illustrating a method for predicting root causes of failures in an exemplary embodiment of this disclosure is provided.

[0043] Figure 10 A block diagram of a training apparatus for a fault root cause prediction model is shown in an exemplary embodiment of this disclosure;

[0044] Figure 11 A block diagram of a fault root cause prediction apparatus is shown in an exemplary embodiment of this disclosure;

[0045] Figure 12 A block diagram illustrating an electronic device according to an exemplary embodiment of the present disclosure. Detailed Implementation

[0046] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0047] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following examples do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0048] It should be noted that the phrase "at least one of several items" in this disclosure refers to three parallel cases: "any one of the several items", "a combination of any number of the several items", and "all of the several items". For example, "including at least one of A and B" includes the following three parallel cases: (1) including A; (2) including B; (3) including A and B. As another example, "performing at least one of step one and step two" indicates the following three parallel cases: (1) performing step one; (2) performing step two; (3) performing both step one and step two.

[0049] Modern industrial manufacturing systems are highly complex, potentially containing thousands of interconnected components. Changes in any component can trigger failures, making it extremely difficult to accurately pinpoint the root cause from a vast array of potential causes (Challenge 1). The complex characteristics of production, such as large scale, nonlinearity, and post-failure non-stationarity, pose significant challenges to root cause diagnosis. Furthermore, the large volume of data, diverse types of data acquisition sensors, severe data redundancy, and low-quality data in production processes limit their ability to express general characteristics (Challenge 2). In addition, complex systems contain numerous interfering factors, which are often not the direct cause of failures but can interfere with fault diagnosis (Challenge 3).

[0050] Therefore, fully exploring the causal relationships of failures and pinpointing the root cause is a very challenging problem. Currently, there are methods that combine traditional knowledge base-based root cause analysis with data-based root cause analysis, such as deep root cause analysis based on knowledge graphs. However, this method only performs root cause reasoning based on the description of the failure phenomenon, ignoring other necessary data influencing factors in actual production data, leading to biased reasoning results. Furthermore, because knowledge bases store a large number of entity relationships with complex interrelationships, this method neglects the analysis and processing of the knowledge base, failing to utilize higher-level information helpful for root cause reasoning.

[0051] If we follow the approach of knowledge graphs and deep neural networks to diagnose root causes of faults, the prediction of root causes primarily relies on semantic similarity for entity linking. Therefore, the importance of semantic feature representation is paramount. Existing methods merely extract semantic features from pre-trained models without further optimizing feature representation. Furthermore, these methods cannot handle scenarios with unknown faults, fail to provide effective guidance for root cause localization, and have poor practicality. Therefore, it is crucial and urgent to improve the effectiveness of root cause analysis and the interpretability of models by analyzing the aforementioned unresolved issues.

[0052] To address the aforementioned issues, this disclosure provides a method for predicting root causes of failures, a model training method, an apparatus, an electronic device, a storage medium, and a computer program product. These methods enable root cause analysis of failures by constructing a hyper-relationship knowledge graph, maximizing the utilization of knowledge and experience information, comprehensively considering various data features, effectively conforming to actual production conditions, and improving the efficiency of root cause prediction.

[0053] Below, we will refer to Figures 1 to 12 This disclosure specifically describes the methods for predicting root causes of failures, model training methods, apparatus, electronic devices, storage media, and computer program products.

[0054] Figure 1 A flowchart illustrating a training method for a fault root cause prediction model in an exemplary embodiment of this disclosure is shown.

[0055] Reference Figure 1 In step 101, training data and a hyperrelationship knowledge graph can be acquired. The training data may include fault cases used for training, each containing at least one fault phenomenon. The hyperrelationship knowledge graph is pre-constructed based on the fault cases and a preset knowledge graph, which contains non-root cause entities and root cause entities. A "root cause entity" can refer to a node in the knowledge graph that represents the root cause of a fault. For example, in a device fault knowledge graph, if a shutdown is caused by "bearing wear," the "bearing wear" node will be marked as a root cause entity. A "non-root cause entity" can refer to other nodes in the knowledge graph that have no direct causal relationship with the fault; that is, a "non-root cause entity" can refer to a node in the same knowledge graph that does not directly participate in the fault causal chain. For example, a normally functioning auxiliary component node within the machine will be marked as a non-root cause entity. The hyperrelationship knowledge graph contains case hyperedges and root cause class hyperedges. Case hyperedges contain non-root cause entities or root cause entities corresponding to each fault phenomenon in the fault case, and root cause class hyperedges contain root cause entities of the same class obtained by clustering.

[0056] According to an exemplary embodiment of this disclosure, a fault case may include multiple fault phenomenon information. For example, the fault phenomenon information may include information such as project classification, import source, equipment type, product, abnormal trigger point, fault phenomenon when it is discovered, private dataset phenomenon category, private dataset fault location, private dataset phenomenon, abnormal content category, first-layer component fault phenomenon, second-layer component fault phenomenon, third-layer component fault phenomenon, project type, detailed project, etc.

[0057] The pre-built knowledge graph can be constructed based on fault cases and other relevant information. Other relevant information may include, but is not limited to: fault case library (e.g., historical maintenance records, experimental data), expert experience (subjective judgment rules of domain experts), literature (e.g., structured knowledge in papers, patents, technical manuals), etc.

[0058] According to an exemplary embodiment of this disclosure, the steps for constructing a hyperrelationship knowledge graph may include, but are not limited to: determining candidate entities corresponding to fault phenomenon information from a preset knowledge graph, wherein the candidate entities are a set of entities distinguished according to ontology classes; determining non-root cause entities or root cause entities corresponding to fault phenomenon information based on the similarity between the text features of the fault phenomenon information and the text features of each entity in the candidate entities corresponding to the fault phenomenon information, so as to obtain case hyperedges of the hyperrelationship knowledge graph; clustering root cause entities in the preset knowledge graph based on the text features of root cause entities, and using the root cause entities of the same clustering class as entities included in the root cause class hyperedges of the hyperrelationship knowledge graph.

[0059] It should be noted that the aforementioned "ontology class" can refer to the category-level nodes that constitute the knowledge graph, which are mainly used for entity classification. "Ontology class differentiation" can refer to the high-level category abstraction of entities.

[0060] According to an exemplary embodiment of this disclosure, during the construction of a hyper-relationship knowledge graph, text features can be extracted from fault phenomenon information, text features can be extracted from entities in a preset knowledge graph, and text features can be extracted from the root causes of faults. Furthermore, the purpose of "extracting text features from the root causes of faults" is to cluster the text features of root cause entities in the preset knowledge graph, and then use the root cause entities of the same cluster as entities contained in the root cause class hyperedges of the hyper-relationship knowledge graph.

[0061] For example, Sentence Bert can be used to extract textual features from the fault phenomenon information of each fault case (hereinafter referred to as a case). N represents the total number of cases, S represents the total number of fault phenomena contained in the cases, and D1 is the feature dimension. Sentence BERT can be used to extract textual features from all non-root cause entities in a pre-defined knowledge graph based on their entity names. M represents the number of all non-root cause entities, and D1 represents the feature dimension. SentenceBERT can be used to extract textual features from the entity name of each root cause entity in the graph. R represents the total number of root cause entities, and D1 represents the feature dimension.

[0062] Next, the factory's fault information can be linked to nodes in a pre-defined knowledge graph. For example, for a fault case, all fault information can be linked to the corresponding nodes in the pre-defined knowledge graph.

[0063] Figure 2 This illustration shows a flowchart of entity linking in an exemplary embodiment of the present disclosure.

[0064] Reference Figure 2 Text features of fault phenomenon information in fault cases can be extracted using Sentence Bert, and text features of root cause entities and non-root cause entities in the preset knowledge graph can be extracted using Sentence Bert. Then, in step (1), candidate entities in the preset knowledge graph can be found for each fault phenomenon information (hereinafter referred to as fault phenomenon) in the fault case.

[0065] Specifically, based on the ontology class corresponding to each entity, we can find the set of all entities contained in each ontology class. This allows us to narrow down the candidate entities for the current fault phenomenon from all entities to a subset of entities, which is the set of entities corresponding to the ontology class. Here, the ontology class is a concept of entity type, and an entity is an instance of the ontology class.

[0066] Then, in step (2), for each fault phenomenon in the fault case, its corresponding entity in the graph can be linked based on the similarity between text features. Specifically, the text features C of the fault phenomenon can be linked... bert With each entity text feature E in the candidate entity set bert Perform a dot product operation to calculate the similarity between the two entities. The entity with the highest similarity can be used as the entity linked to the fault phenomenon.

[0067] Steps (1) and (2) above can be repeated to link entities for each case. Let each case contain S entities, and there are a total of N cases.

[0068] Next, text clustering can be performed on the root causes of the failure.

[0069] Figure 3 This illustration shows a flowchart of text clustering for root causes of failures in an exemplary embodiment of the present disclosure.

[0070] Reference Figure 3Given a pre-defined knowledge graph containing R root cause entities, after extracting their textual features using Sentence BERT, Supporting Clustering with Contrastive Learning (SCCL) can be used to cluster these R root cause entities, grouping semantically similar root causes into one category. K cluster centers can be determined using the elbow method and silhouette coefficient method. This allows us to obtain the root cause entities contained in each cluster center.

[0071] Hyperrelational knowledge graphs are composed of hyperrelational facts (H-Facts). An H-Fact contains a basic triple and a series of related attribute-value pairs, represented as follows: Where (s,r,o) represents the basic triplet, (v i :e i () represents attribute-value pairs, V is the entity set, and E is the relation set. Since H-Facts contain more than two entities, this disclosure introduces the concept of a hypergraph. H-Facts are considered as hyperedges, and entities in the graph are considered as nodes. Each hyperedge contains multiple nodes (greater than or equal to 2).

[0072] Figure 4 This illustration shows a flowchart of constructing a hyperrelationship knowledge graph in an exemplary embodiment of this disclosure.

[0073] Reference Figure 4 We can define case hyperedges and root cause class hyperedges separately. Specifically, a case can be considered a hyperedge, containing S entities, for a total of N case hyperedges; a root cause cluster center can be considered a hyperedge, with the root causes belonging to that cluster being entities contained within the hyperedge, for a total of K root cause class hyperedges. Finally, the super-relationship knowledge graph with cases as hyperedges and the super-relationship knowledge graph with root cause classes as hyperedges can be merged into a single graph, forming a super-relationship knowledge graph.

[0074] Return to reference Figure 1 In step 102, the fault cases and the super-relationship knowledge graph are input into the fault root cause prediction model to obtain the predicted value of the fault root cause.

[0075] According to an exemplary embodiment of this disclosure, a fault root cause prediction model includes multiple learnable parameter matrices. The process of inputting fault cases and a hyperrelationship knowledge graph into the fault root cause prediction model to obtain predicted values ​​of fault root causes includes: inputting fault cases and a hyperrelationship knowledge graph into the fault root cause prediction model, and performing the following operations through the fault root cause prediction model: extracting global case features of case hyperedges and global root cause class features of root cause class hyperedges using a first learnable parameter matrix and a second learnable parameter matrix, respectively; and obtaining a prediction of the root cause class of the fault root cause based on the dot product of the global case features and the global root cause class features. The process involves: filtering global case features based on the third learnable parameter matrix to obtain local case features; filtering global root cause class features based on the fourth learnable parameter matrix to obtain local root cause features; adjusting the weights of corresponding root causes in the local root cause features based on the predicted values ​​of the root cause classes to obtain weight-adjusted local root cause features; and obtaining the predicted value of the fault root cause based on the dot product between the local case features and the weight-adjusted local root cause features. The training of the fault root cause prediction model based on the first loss includes: training multiple learnable parameter matrices in the fault root cause prediction model based on the first loss.

[0076] According to exemplary embodiments of this disclosure, coarse-grained root cause prediction and fine-grained root cause prediction can be performed based on a hyper-relationship knowledge graph.

[0077] Specifically, we can first extract the structural features of entities in the hyperrelationship knowledge graph.

[0078] For example, HAHE (Hierarchical Attention Heterogeneous Information Network Embedding) can be used to extract the structural features of entities in hyperrelationship knowledge graphs. (M+R) represents the total number of entities in the hyper-relationship knowledge graph, and D² represents the feature dimension. HAHE can learn the features of the hyper-relationship knowledge graph through both global and local attention mechanisms. The aggregation and update process from nodes to hyperedges can be implemented using an attention mechanism.

[0079]

[0080] in, Indicates the superedge e j eigenvectors, Represents node v i For the superedge e j The importance of , where · denotes the dot product operation. δ and σ are both ReLU activation functions. W and W vThis is a learnable parameter matrix. The aggregation and update process from hyperedges to nodes is the same as above, which yields the node features.

[0081] According to an exemplary embodiment of this disclosure, global case features of case hyperedges and global root cause features of root cause hyperedges are extracted using a first learnable parameter matrix and a second learnable parameter matrix, respectively. This may include, but is not limited to, the following steps: extracting text features of fault cases and structural features within case hyperedges, and combining the extracted text features and structural features with the first learnable parameter matrix to obtain global case features of case hyperedges; extracting text features and structural features within root cause hyperedges, and combining the extracted text features and structural features with the second learnable parameter matrix to obtain global root cause features of root cause hyperedges.

[0082] According to exemplary embodiments of this disclosure, global-level case features and root cause features, i.e., global case features and global root cause features, can be obtained.

[0083] Specifically, global case features can be obtained through the following steps:

[0084] Within a case hyperedge, the structural features E of the associated node can be... struct Mean-based aggregation yields the feature embeddings of the case hyperedges. N is the number of cases, and D3 is the feature dimension. A hyperedge e j The calculation formula can be as follows: num(e j ) represents the superedge e j Number of nodes included:

[0085]

[0086] Where l represents the number of network layers.

[0087] Next, the textual features C of the case can be... bert and structural features C struct The data is mapped to the same spatial context and then concatenated. After passing through a fully connected layer, the global case features, i.e., the fused case features, are obtained. The calculation formula can be as follows:

[0088]

[0089] Among them, W a W b W t and b a All of these are learnable parameter matrices, which can be denoted as the first learnable parameter matrix, and σ is the ReLU activation function.

[0090] Global root cause features can be obtained through the following steps:

[0091] Within a root cause class hyperedge, the structural features E of the associated node can be... struct Mean-based aggregation, using the same aggregation method as above, yields the features of root cause class hyperedges. K is the total number of cluster centers, and D3 is the feature dimension.

[0092] Next, the text features A of the root cause class can be... bert and structural feature A struct The features are mapped to the same spatial dimension and concatenated, then passed through a fully connected layer to obtain the global root cause features, i.e., the fused root cause features. The calculation formula can be as follows:

[0093]

[0094] Among them, W c W d W r and b c Both are learnable parameter matrices, which can be denoted as the second learnable parameter matrix, and σ is the activation function ReLU.

[0095] According to exemplary embodiments of this disclosure, coarse-grained root cause prediction can be performed based on a hyper-relationship knowledge graph. Specifically, the obtained global case features can be... and global root cause features Perform a dot product to obtain the predicted result Y. 1 That is, the predicted value of the root cause class of the root cause of the failure.

[0096] According to exemplary embodiments of this disclosure, local-level case features and root cause features, i.e., local case features and local root cause features, can be obtained.

[0097] Specifically, local case features can be obtained through the following steps:

[0098] Global case features can be used By using a self-gate filter to encode higher-order information, local case features are obtained. The self-gate is:

[0099]

[0100] Among them, W s and b s It is a learnable parameter matrix, which can be denoted as the third learnable parameter matrix, and * is the element-wise product.

[0101] According to an exemplary embodiment of this disclosure, filtering global root cause features based on a fourth learnable parameter matrix to obtain local root cause features may include, but is not limited to, the following steps: expanding the global root cause features corresponding to the predicted values ​​of root cause classes based on the number of root causes in the hyper-relationship knowledge graph to obtain first global root cause features; obtaining first structural features of root cause entities in the predicted values ​​of root cause classes based on the index of root causes in the hyper-relationship knowledge graph; and concatenating and filtering the first global root cause features and the first structural features based on the fourth learnable parameter matrix to obtain local root cause features.

[0102] According to exemplary embodiments of this disclosure, local root cause features can be obtained.

[0103] Based on the root cause class predicted by the above steps, the root cause class of each case can be determined, meaning that the root cause of the case is very likely to be the currently predicted root cause class.

[0104] Specifically, based on the predicted k-th root cause class, where k is the index of the predicted root cause class relative to all root cause classes, the global root cause class features can be taken. And Expanding R times yields a size of R*D4. Since there are R root causes, the structural features of the root cause nodes can be obtained based on their entity numbers (indices).

[0105] Next, the two features can be concatenated and then passed through a self-gate filter to encode higher-order root cause information, resulting in fused local root cause features. The formula is:

[0106]

[0107] Among them, W u W k W e b e All are learnable parameter matrices, which can be denoted as the fourth learnable parameter matrix. ⊙ is the element-wise product, and u is the number of root causes in the predicted root cause class.

[0108] Since the predicted root causes within a given root cause class are highly likely to be correct, it is desirable for the model to focus more on these root causes. Therefore, this disclosure designs a specific attention mechanism to direct the model's attention to these nodes. Based on the k-th root cause class predicted in the above steps, it is known that the u root causes within this class are more likely to be true root causes. Therefore, the attention association matrix T∈R, initialized with all elements set to 1, can be used. R *RIn this context, the indices corresponding to the u root causes are multiplied by a larger weight α, while the remaining root causes are multiplied by a smaller weight (1-α), where the value of α is greater than 0.5.

[0109] Finally, the correlation matrix T can be combined with the fused local root cause features A. + Multiplying by elements yields the weighted local root cause features.

[0110] A * =T⊙A +

[0111] According to exemplary embodiments of this disclosure, fine-grained root cause prediction can be performed based on a hyperrelationship knowledge graph.

[0112] Figure 5 This illustration shows a flowchart of fine-grained root cause prediction based on a hyperrelationship knowledge graph in an exemplary embodiment of this disclosure.

[0113] Reference Figure 5 According to an exemplary embodiment of this disclosure, local case features are obtained by filtering global case features. Local root cause features are obtained by combining global root cause features with a candidate root cause set (i.e., predicted root cause classes) through feature concatenation, filtering, and a special attention mechanism. Local case features and local root cause features can be used to obtain a local hyperrelationship knowledge graph. The obtained local case features C... * and local root cause characteristics A * The dot product can predict the possible root cause Y for each case. 2 That is, the predicted value of the root cause of the failure.

[0114] Combined with reference Figure 1 In step 103, the first loss between the predicted value and the corresponding true value of the root cause of the fault is calculated.

[0115] The predicted value Y of the root cause of the failure can be calculated using the cross-entropy formula. 2 Compared with the true value Loss 1, or the first loss:

[0116] Y 2 =C * *A* T

[0117]

[0118] In step 104, the fault root cause prediction model is trained based on the first loss.

[0119] According to an exemplary embodiment of this disclosure, multiple learnable parameter matrices, including a first learnable parameter matrix, a second learnable parameter matrix, a third learnable parameter matrix, a fourth learnable parameter matrix, etc., in the fault root cause prediction model can be adjusted based on a first loss to train the fault root cause prediction model.

[0120] According to an exemplary embodiment of this disclosure, the training method for the fault root cause prediction model further includes, but is not limited to, the following steps: calculating a second loss between the predicted value and the corresponding true value of the root cause class; wherein, training the fault root cause prediction model based on the first loss includes: training a plurality of learnable parameter matrices in the fault root cause prediction model based on the first loss and the second loss.

[0121] According to an exemplary embodiment of this disclosure, the predicted value Y of the root cause class of the root cause of the fault can be calculated using the cross-entropy formula. 1 Compared with the true value Loss2, or the second loss:

[0122]

[0123] Figure 6 This illustration shows a flowchart of coarse-grained root cause prediction based on a hyperrelationship knowledge graph in an exemplary embodiment of this disclosure.

[0124] Reference Figure 6 According to an exemplary embodiment of this disclosure, for a global-level hyperrelation knowledge graph containing case hyperedges and root cause hyperedges, textual features of cases and textual features of root causes can be extracted from it. Structural features of case hyperedges and structural features of root cause hyperedges can be extracted using HAHE, respectively, and then fused with the textual features of cases and the textual features of root causes to obtain global case features and global root cause features. Based on the global case features and global root cause features, the predicted value of the root cause class of the fault root cause can be obtained, and a second loss can be calculated between the predicted value and the true value of the root cause class.

[0125] Multiple learnable parameter matrices in the fault root cause prediction model can be trained by combining the first loss and the second loss.

[0126] According to an exemplary embodiment of this disclosure, the training method of the fault root cause prediction model further includes, but is not limited to, the following steps: calculating a third loss through contrastive learning based on the similarity between the text features of fault cases and the text features of root cause entities in the hyperrelationship knowledge graph; training the fault root cause prediction model based on the first loss may include, but is not limited to, the following steps: training multiple learnable parameter matrices in the fault root cause prediction model based on the first loss and the third loss.

[0127] According to exemplary embodiments of this disclosure, feature representation optimization can be achieved through contrastive learning.

[0128] Specifically, the similarity between the text features of the case and the text features of the root cause can be calculated. Since the word combinations of the fault symptoms and the meanings expressed by the root cause words are very similar, contrastive learning can be used to maximize the text features describing the fault symptoms in the case. Root cause text features The mutual information between them. The two features mentioned above can be unified into the same space through a contrastive learning head. Specifically, the contrastive learning head network sequence is a fully connected layer, a ReLU activation function, and another fully connected layer. Then, the third loss can be calculated using contrastive learning.

[0129]

[0130] Where sim(·,·) is the dot product operation after the features have been normalized, τ is the temperature coefficient, set to 0.4, and c bert i a is the text feature of the i-th case. bert i It is the textual feature of the true root cause corresponding to the i-th case, (c bert i ,a bert i ) constitute positive sample pairs, a bert j Other root causes within all batches, (c bert i ,a bert j ) j≠i This forms a negative sample pair.

[0131] Multiple learnable parameter matrices in the fault root cause prediction model can be trained by combining the first loss and the third loss.

[0132] According to exemplary embodiments of this disclosure, multiple learnable parameter matrices in a fault root cause prediction model can be trained based on a first loss, a second loss, and a third loss.

[0133] Specifically, the tasks of root cause fuzzy prediction (predicting root cause classes), root cause correction prediction (weight adjustment), and self-supervised (contrastive learning) can be unified into a master-slave learning framework, where the first two are the master tasks and the latter is the auxiliary task. The joint learning loss (Loss) can be defined as:

[0134] Loss = Loss1 + βLoss2 + γLoss3

[0135] Here, β and γ are hyperparameters.

[0136] Figure 7 This diagram illustrates a flowchart of a method for training a root cause prediction model in an exemplary embodiment of the present disclosure.

[0137] Reference Figure 7 According to exemplary embodiments of this disclosure, for a global-level hyperrelation knowledge graph containing case hyperedges and root cause hyperedges, textual features of cases and textual features of root causes can be extracted. The similarity between the two can be calculated using a contrastive learning head to calculate a third loss. Structural feature extraction of case hyperedges and structural feature extraction of root cause hyperedges can be performed using HAHE, respectively, and fused with the textual features of cases and root causes to obtain global case features and global root cause features. Based on the global case features and global root cause features, a predicted value of the root cause class of the fault root cause can be obtained, and further calculated... The second loss between the predicted and true values ​​of the root cause class; after filtering the global case features, local case features are obtained; the global root cause class features are combined with the candidate root cause set (i.e., the predicted root cause class) and then subjected to feature concatenation, filtering, and a special attention mechanism to obtain local root cause features; local case features and local root cause features can be used to obtain a local hyperrelationship knowledge graph; based on the local case features and local root cause features, the predicted value of the fault root cause for each case can be obtained, and the first loss between the predicted and true values ​​of the fault root cause can be calculated; multiple learnable parameter matrices in the fault root cause prediction model can be trained by combining the first loss, the second loss, and the third loss.

[0138] According to exemplary embodiments of the present disclosure, the training method of the fault root cause prediction model of the present disclosure can mainly include three modules: first, super-relationship knowledge graph construction; second, coarse-grained root cause prediction and fine-grained root cause prediction based on super-relationship knowledge graph; and third, feature representation optimization based on contrastive learning.

[0139] Figure 8 This illustration shows a flowchart of a self-supervised fault root cause analysis framework based on a hyperrelationship knowledge graph in an exemplary embodiment of this disclosure.

[0140] Reference Figure 8 According to an exemplary embodiment of this disclosure, the process can be mainly divided into three parts: super-relationship knowledge graph construction, multi-granularity root cause analysis network, and feature representation optimization.

[0141] In the process of constructing the hyperrelationship knowledge graph, case hyperedges are constructed by linking entities to the failure phenomena of historical cases; root cause class hyperedges are constructed by clustering texts to identify the root causes of failures.

[0142] A multi-granularity root cause analysis network based on a hyper-relationship knowledge graph can serve as the main task in a master-slave framework, which can include two steps:

[0143] Step 1: Predict candidate root causes, i.e. root cause classes, based on a coarse-grained root cause analysis network of super-relational knowledge graph. Here, a loss function loss2, i.e., the second loss function, can be designed.

[0144] Step 2: The fine-grained root cause analysis network based on the super-relationship knowledge graph adds weights to the candidate root causes and predicts the root causes in the full range. Here, a loss function loss1, i.e. the first loss function, can be designed.

[0145] For feature representation optimization, it can be used as an auxiliary task in the main-auxiliary framework to optimize feature representation based on semantic similarity constraints of contrastive learning. Here, a loss function loss3, i.e., the third loss function, can be designed.

[0146] According to exemplary embodiments of the present disclosure, the fault root cause prediction model trained by the training method of the present disclosure has strong domain versatility and is applicable to various factory faults and automated production systems, etc.; and the training method of the fault root cause prediction model of the present disclosure is flexible and can flexibly adjust each module according to different working conditions, such as the construction of a hyper-relationship knowledge graph, etc.

[0147] The exemplary embodiments disclosed herein can be applied to large-scale fault root cause analysis tasks in semiconductor factories, improving production efficiency, reducing costs and increasing efficiency, and assisting experienced experts in making decisions on fault root cause analysis. Furthermore, it is highly versatile and can be extended to fault root cause analysis of other large-scale factory equipment, such as automated production systems like motor repair, automobile manufacturing, and blast furnace and roll control systems in metallurgy. The modules of the scheme can be flexibly adjusted according to the number of different fault phenomena, for example, increasing the number of ontology classes during the hyperrelationship knowledge graph modeling process.

[0148] Figure 9 A flowchart illustrating a method for predicting root causes of failures in an exemplary embodiment of this disclosure is shown.

[0149] Reference Figure 9 In step 901, a fault description and a super-relationship knowledge graph are obtained. The fault description contains at least one fault phenomenon information. The super-relationship knowledge graph is pre-constructed based on fault cases and a preset knowledge graph. The fault case contains at least one fault phenomenon information. The preset knowledge graph contains non-root cause entities and root cause entities. The super-relationship knowledge graph contains case super edges and root cause class super edges. The case super edge contains the non-root cause entity or root cause entity corresponding to each fault phenomenon information in the fault case. The root cause class super edge contains root cause entities of the same class obtained by clustering.

[0150] According to an exemplary embodiment of this disclosure, the root cause prediction model includes multiple learnable parameter matrices. The model inputs a fault description and a hyperrelationship knowledge graph to obtain predicted values ​​for the root causes of the fault, including but not limited to the following steps: The root cause prediction model performs the following operations: extracting global case features of case hyperedges and global root cause class features of root cause class hyperedges using a first and a second learnable parameter matrix, respectively; obtaining the predicted value of the root cause class based on the dot product of the global case features and the global root cause class features; obtaining local case features by filtering the global case features using a third learnable parameter matrix; obtaining local root cause features by filtering the global root cause class features using a fourth learnable parameter matrix; adjusting the weights of the corresponding root causes in the local root cause features based on the predicted values ​​of the root cause classes to obtain weight-adjusted local root cause features; and obtaining the predicted value of the root cause based on the dot product between the local case features and the weight-adjusted local root cause features.

[0151] According to an exemplary embodiment of this disclosure, filtering global root cause features based on a fourth learnable parameter matrix to obtain local root cause features may include, but is not limited to, the following steps: expanding the global root cause features corresponding to the predicted values ​​of root cause classes based on the number of root causes in the hyper-relationship knowledge graph to obtain first global root cause features; obtaining first structural features of root cause entities in the predicted values ​​of root cause classes based on the index of root causes in the hyper-relationship knowledge graph; and concatenating and filtering the first global root cause features and the first structural features based on the fourth learnable parameter matrix to obtain local root cause features.

[0152] According to an exemplary embodiment of this disclosure, global case features of case hyperedges and global root cause features of root cause hyperedges are extracted using a first learnable parameter matrix and a second learnable parameter matrix, respectively. This may include, but is not limited to, the following steps: extracting text features of fault cases and structural features within case hyperedges, and combining the extracted text features and structural features with the first learnable parameter matrix to obtain global case features of case hyperedges; extracting text features and structural features within root cause hyperedges, and combining the extracted text features and structural features with the second learnable parameter matrix to obtain global root cause features of root cause hyperedges.

[0153] According to an exemplary embodiment of this disclosure, a hyperrelationship knowledge graph can be constructed through the following steps: determining candidate entities corresponding to fault phenomenon information from a preset knowledge graph, wherein the candidate entities are a set of entities distinguished according to ontology classes; determining non-root cause entities or root cause entities corresponding to fault phenomenon information based on the similarity between the text features of the fault phenomenon information and the text features of each entity in the candidate entities corresponding to the fault phenomenon information, so as to obtain case hyperedges of the hyperrelationship knowledge graph; clustering based on the text features of root cause entities in the preset knowledge graph, and taking the root cause entities of the same class obtained by clustering as the entities contained in the root cause class hyperedges of the hyperrelationship knowledge graph.

[0154] In step 902, the fault description and hyperrelationship knowledge graph are input into the fault root cause prediction model to obtain the predicted value of the fault root cause; wherein, the fault root cause prediction model is trained according to the above-mentioned fault root cause prediction model training method.

[0155] It is understood that the specific implementation process of the above-described root cause prediction method is largely the same as that of the above-described root cause prediction model training method, and will not be elaborated here.

[0156] Figure 10 A block diagram of a training apparatus for a fault root cause prediction model is shown in an exemplary embodiment of this disclosure.

[0157] Reference Figure 10 The exemplary embodiments of this disclosure also provide a training apparatus 1000 for a fault root cause prediction model, which may include, but is not limited to, a data acquisition unit 1001, a root cause prediction unit 1002, a first loss calculation unit 1003, and a model training unit 1004.

[0158] The data acquisition unit 1001 can acquire training data and a hyper-relationship knowledge graph. The training data may include fault cases used for training, each fault case containing at least one fault phenomenon information. The hyper-relationship knowledge graph is pre-constructed based on the fault cases and a preset knowledge graph. The preset knowledge graph contains non-root cause entities and root cause entities. The hyper-relationship knowledge graph contains case hyperedges and root cause class hyperedges. The case hyperedges contain non-root cause entities or root cause entities corresponding to each fault phenomenon information in the fault case. The root cause class hyperedges contain root cause entities of the same class obtained by clustering.

[0159] The root cause prediction unit 1002 can input fault cases and hyperrelationship knowledge graphs into the fault root cause prediction model to obtain the predicted value of the fault root cause.

[0160] The first loss calculation unit 1003 can calculate the first loss between the predicted value and the corresponding true value of the root cause of the fault.

[0161] Model training unit 1004 can train the fault root cause prediction model based on the first loss.

[0162] According to an exemplary embodiment of this disclosure, the root cause prediction model includes multiple learnable parameter matrices. The root cause prediction unit 1002 can input fault cases and a hyperrelationship knowledge graph into the root cause prediction model, and perform the following operations: extracting global case features of case hyperedges and global root cause class features of root cause class hyperedges using a first and a second learnable parameter matrix, respectively; obtaining the predicted value of the root cause class of the fault root cause based on the dot product of the global case features and the global root cause class features; obtaining local case features by filtering the global case features based on a third learnable parameter matrix; obtaining local root cause features by filtering the global root cause class features based on a fourth learnable parameter matrix; adjusting the weights of the corresponding root causes in the local root cause features based on the predicted value of the root cause class, obtaining weight-adjusted local root cause features; and obtaining the predicted value of the fault root cause based on the dot product between the local case features and the weight-adjusted local root cause features. The model training unit 1004 can train the multiple learnable parameter matrices in the fault root cause prediction model based on a first loss.

[0163] According to an exemplary embodiment of the present disclosure, the training apparatus 1000 for the root cause prediction model may further include a second loss calculation unit (not shown) that can calculate a second loss between the predicted value of the root cause class and the corresponding true value; wherein, the model training unit 1004 can train a plurality of learnable parameter matrices in the root cause prediction model based on the first loss and the second loss.

[0164] According to an exemplary embodiment of the present disclosure, the training apparatus 1000 for the fault root cause prediction model may further include a third loss calculation unit (not shown), which can calculate the third loss through comparative learning based on the similarity between the text features of the fault case and the text features of the root cause entity in the hyperrelationship knowledge graph; the model training unit 1004 can train multiple learnable parameter matrices in the fault root cause prediction model based on the first loss and the third loss.

[0165] According to an exemplary embodiment of the present disclosure, the training apparatus 1000 for the fault root cause prediction model may further include a third loss calculation unit (not shown), which can calculate the third loss through contrastive learning based on the similarity between the text features of the fault case and the text features of the root cause entity in the hyperrelationship knowledge graph; wherein, the model training unit 1004 can train multiple learnable parameter matrices in the fault root cause prediction model based on the first loss, the second loss and the third loss.

[0166] It is understood that the specific implementation process of the training device 1000 for the fault root cause prediction model described above is largely the same as that of the training method for the fault root cause prediction model described above, and will not be repeated here. The training device 1000 for the fault root cause prediction model can be configured as software, hardware, firmware, or any combination thereof to perform specific functions. For example, these devices may correspond to dedicated integrated circuits, pure software code, or modules combining software and hardware. Furthermore, one or more functions implemented by these devices may also be uniformly executed by components in a physical entity device (e.g., a processor, client, or server).

[0167] Figure 11 A block diagram of a fault root cause prediction apparatus is shown in an exemplary embodiment of this disclosure.

[0168] Reference Figure 11 An exemplary embodiment of this disclosure also provides a fault root cause prediction device 1100, which may include, but is not limited to, a data acquisition unit 1101 and a root cause prediction unit 1102.

[0169] The data acquisition unit 1101 can acquire fault descriptions and super-relationship knowledge graphs. The fault descriptions contain at least one fault phenomenon information. The super-relationship knowledge graphs are pre-constructed based on fault cases and preset knowledge graphs. Fault cases contain at least one fault phenomenon information. The preset knowledge graphs contain non-root cause entities and root cause entities. The super-relationship knowledge graphs contain case hyperedges and root cause class hyperedges. Case hyperedges contain non-root cause entities or root cause entities corresponding to each fault phenomenon information in the fault case. Root cause class hyperedges contain root cause entities of the same class obtained by clustering.

[0170] The root cause prediction unit 1102 can input the fault description and hyperrelationship knowledge graph into the fault root cause prediction model to obtain the predicted value of the fault root cause; wherein, the fault root cause prediction model is trained according to the above-mentioned fault root cause prediction model training method.

[0171] According to an exemplary embodiment of this disclosure, the root cause prediction model includes multiple learnable parameter matrices. The root cause prediction unit 1102 can input fault descriptions and hyperrelationship knowledge graphs into the root cause prediction model, and perform the following operations: extracting global case features of case hyperedges and global root cause class features of root cause class hyperedges using a first and a second learnable parameter matrix, respectively; obtaining the predicted value of the root cause class of the fault root cause based on the dot product of the global case features and the global root cause class features; obtaining local case features by filtering the global case features based on a third learnable parameter matrix; obtaining local root cause features by filtering the global root cause class features based on a fourth learnable parameter matrix; adjusting the weights of the corresponding root causes in the local root cause features based on the predicted value of the root cause class, obtaining the weight-adjusted local root cause features; and obtaining the predicted value of the fault root cause based on the dot product between the local case features and the weight-adjusted local root cause features.

[0172] According to an exemplary embodiment of this disclosure, the root cause prediction unit 1102 can expand the global root cause class features corresponding to the predicted value of the root cause class based on the number of root causes in the super-relationship knowledge graph to obtain the first global root cause class features; obtain the first structural features of the root cause entity in the predicted value of the root cause class based on the index of the root cause in the super-relationship knowledge graph; and concatenate and filter the first global root cause class features and the first structural features based on the fourth learnable parameter matrix to obtain local root cause features.

[0173] According to an exemplary embodiment of this disclosure, a hyperrelationship knowledge graph is constructed through the following steps: determining candidate entities corresponding to fault phenomenon information from a preset knowledge graph, wherein the candidate entities are a set of entities distinguished according to ontology classes; determining non-root cause entities or root cause entities corresponding to fault phenomenon information based on the similarity between the text features of the fault phenomenon information and the text features of each entity in the candidate entities corresponding to the fault phenomenon information, so as to obtain case hyperedges of the hyperrelationship knowledge graph; clustering based on the text features of root cause entities in the preset knowledge graph, and taking the root cause entities of the same class obtained by clustering as the entities contained in the root cause class hyperedges of the hyperrelationship knowledge graph.

[0174] According to an exemplary embodiment of this disclosure, the root cause prediction unit 1102 can extract textual features of the fault description and structural features within the case hyperedge, and obtain global case features of the case hyperedge based on the extracted textual features and structural features combined with a first learnable parameter matrix; and obtain global root cause class features of the root cause class hyperedge by extracting textual features and structural features within the root cause class hyperedge, and obtaining global root cause class features of the root cause class hyperedge based on the extracted textual features and structural features combined with a second learnable parameter matrix.

[0175] It is understood that the specific implementation process of the above-described root cause prediction device 1100 is largely the same as that of the above-described root cause prediction method, and will not be described in detail here. The root cause prediction device 1100 can be configured as software, hardware, firmware, or any combination thereof to perform specific functions. For example, these devices may correspond to dedicated integrated circuits, pure software code, or modules combining software and hardware. Furthermore, one or more functions implemented by these devices may also be uniformly executed by components in a physical entity device (e.g., a processor, client, or server).

[0176] Figure 12 A block diagram illustrating an electronic device according to an exemplary embodiment of the present disclosure.

[0177] Reference Figure 12 The electronic device 1200 includes at least one memory 1201 and at least one processor 1202. The at least one memory 1201 stores a set of computer-executable instructions. When the set of computer-executable instructions is executed by the at least one processor 1202, a training method for a fault root cause prediction model and a fault root cause prediction method according to exemplary embodiments of the present disclosure are executed.

[0178] As an example, electronic device 1200 may be a PC, tablet, personal digital assistant, smartphone, or other device capable of executing the aforementioned set of instructions. Here, electronic device 1200 is not necessarily a single electronic device, but may be any collection of devices or circuits capable of executing the aforementioned instructions (or instruction sets) individually or in combination. Electronic device 1200 may also be part of an integrated control system or system manager, or may be configured to interconnect with a portable electronic device locally or remotely (e.g., via wireless transmission) through an interface.

[0179] In electronic device 1200, processor 1202 may include a central processing unit (CPU), a graphics processing unit (GPU), a programmable logic device, a dedicated processor system, a microcontroller, or a microprocessor. By way of example and not limitation, processor may also include analog processors, digital processors, microprocessors, multi-core processors, processor arrays, network processors, etc.

[0180] The processor 1202 can execute instructions or code stored in the memory 1201, which can also store data. Instructions and data can also be sent and received over a network via a network interface device, which can employ any known transmission protocol.

[0181] The memory 1201 may be integrated with the processor 1202, for example, by arranging RAM or flash memory within an integrated circuit microprocessor. Alternatively, the memory 1201 may include a separate device, such as an external disk drive, a storage array, or other storage device usable by any database system. The memory 1201 and the processor 1202 may be operatively coupled, or may communicate with each other, for example, via I / O ports, network connections, etc., enabling the processor 1202 to read files stored in the memory.

[0182] In addition, the electronic device 1200 may also include a video display (such as a liquid crystal display) and a user interaction interface (such as a keyboard, mouse, touch input device, etc.). All components of the electronic device 1200 can be interconnected via a bus and / or network.

[0183] According to exemplary embodiments of the present disclosure, a computer-readable storage medium storing instructions may also be provided, wherein when the instructions are executed by at least one computing device, they cause at least one computing device to perform the training method of the fault root cause prediction model and the fault root cause prediction method described above.

[0184] Examples of computer-readable storage media herein include: read-only memory (ROM), random access programmable read-only memory (PROM), electrically erasable programmable read-only memory (EEPROM), random access memory (RAM), dynamic random access memory (DRAM), static random access memory (SRAM), flash memory, non-volatile memory, CD-ROM, CD-R, CD+R, CD-RW, CD+RW, DVD-ROM, DVD-R, DVD+R, DVD-RW, DVD+RW, DVD-RAM, BD-ROM, BD-R, BD-R LTH, BD-RE, Blu-ray or optical disc storage, hard disk drive (HDD), solid-state drive (SSD), card storage (such as multimedia cards, secure digital (SD) cards, or ultra-fast digital (XD) cards), magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, and any other device configured to store a computer program and any associated data, data files, and data structures in a non-transitory manner and to provide the computer program and any associated data, data files, and data structures to a processor or computer so that the processor or computer can execute the computer program. The computer program in the aforementioned computer-readable storage medium can run in an environment deployed in computer devices such as clients, hosts, agent devices, servers, etc. Furthermore, in one example, the computer program and any associated data, data files, and data structures are distributed across a networked computer system, such that the computer program and any associated data, data files, and data structures are stored, accessed, and executed in a distributed manner through one or more processors or computers. It should be noted that the instructions can also be used to perform additional steps in addition to those described above, or to perform more specific processing while performing the above steps. The details of these additional steps and further processing have already been mentioned in the description of the relevant methods, so they will not be repeated here to avoid repetition.

[0185] Another embodiment of this disclosure relates to a system comprising a storage device including at least one computing device and at least one stored instruction, wherein the instruction, when executed by at least one computing device, causes at least one computing device to perform the above-described training method for the root cause prediction model and the root cause prediction method.

[0186] It should be noted that the system according to the exemplary embodiments of this disclosure may rely entirely on the operation of computer programs or instructions to achieve the corresponding functions. That is, each unit corresponds to each step in the functional architecture of the computer program, so that the entire system is called through a special software package (e.g., a lib library) to achieve the corresponding functions.

[0187] On the other hand, when the above system is implemented in software, firmware, middleware or microcode, the program code or code segment used to perform the corresponding operation can be stored in a computer-readable medium such as a storage medium, so that at least one processor or at least one computing device can perform the corresponding operation by reading and running the corresponding program code or code segment.

[0188] According to exemplary embodiments of this disclosure, the storage device may be integrated with the computing device, for example, by arranging RAM or flash memory within an integrated circuit microprocessor. Alternatively, the storage device may include a separate device, such as an external disk drive, a storage array, or other storage device usable by any database system. The storage device and the computing device may be operatively coupled, or may communicate with each other, for example, via I / O ports, network connections, etc., enabling the computing device to read instructions stored in the storage device.

[0189] Another embodiment of this disclosure relates to a computer program product, including a computer program / instructions that, when executed by a processor, implement the training method for the fault root cause prediction model and the fault root cause prediction method as described in any of the above.

[0190] Based on the failure root cause prediction method, model training method, device, electronic equipment, storage medium and computer program product provided in this disclosure, failure root cause analysis can be performed by constructing a hyper-relationship knowledge graph, thereby maximizing the utilization of knowledge and experience information, realizing comprehensive consideration of multiple data features, truly conforming to actual production conditions, and improving the efficiency of root cause prediction.

[0191] Furthermore, combining coarse-grained and fine-grained root cause prediction can improve the interpretability of knowledge-based deep learning models, enabling their application in actual production.

[0192] In addition, comparative learning can optimize the semantic feature representation of fault information, thereby helping to improve the effectiveness of the main task of fault root cause analysis.

[0193] The foregoing has described various exemplary embodiments of this disclosure. It should be understood that the foregoing description is exemplary only and not exhaustive, and this disclosure is not limited to the disclosed exemplary embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of this disclosure. Therefore, the scope of protection of this disclosure should be determined by the scope of the claims.

Claims

1. A method for predicting the root causes of failures, characterized in that, The method for predicting the root causes of failures includes: Obtain a fault description and a super-relationship knowledge graph, wherein the fault description contains at least one fault phenomenon information, the super-relationship knowledge graph is pre-constructed based on fault cases and a preset knowledge graph, the fault case contains at least one fault phenomenon information, the preset knowledge graph contains non-root cause entities and root cause entities, the super-relationship knowledge graph contains case super edges and root cause class super edges, the case super edge contains the non-root cause entity or the root cause entity corresponding to each fault phenomenon information in the fault case, and the root cause class super edge contains the root cause entities of the same class obtained by clustering; The fault description and the hyperrelationship knowledge graph are input into the fault root cause prediction model to obtain the predicted value of the fault root cause.

2. The method for predicting the root cause of a failure as described in claim 1, characterized in that, The root cause prediction model includes multiple learnable parameter matrices. The step of inputting the fault description and the hyperrelationship knowledge graph into the fault root cause prediction model to obtain the predicted value of the fault root cause includes: The following operations are performed using the aforementioned root cause prediction model: Using the first learnable parameter matrix and the second learnable parameter matrix, the global case features of the case hyperedge and the global root cause class features of the root cause class hyperedge are extracted respectively. Based on the dot product of the global case features and the global root cause class features, the predicted value of the root cause class of the fault root cause is obtained. Local case features are obtained by filtering the global case features based on the third learnable parameter matrix; Based on the filtering of the global root cause class features using the fourth learnable parameter matrix, local root cause features are obtained. Based on the predicted value of the root cause class, the weights of the corresponding root causes in the local root cause features are adjusted to obtain the weight-adjusted local root cause features. The predicted value of the fault root cause is obtained based on the dot product between the local case features and the weighted local root cause features.

3. The method for predicting the root cause of a failure as described in claim 2, characterized in that, The filtering of the global root cause features based on the fourth learnable parameter matrix to obtain local root cause features includes: Based on the number of root causes in the super-relationship knowledge graph, the global root cause class features corresponding to the predicted values ​​of the root cause class are expanded to obtain the first global root cause class features. Based on the index of the root cause in the super-relationship knowledge graph, the first structural feature of the root cause entity in the predicted value of the root cause class is obtained. The first global root cause features and the first structural features are concatenated and filtered based on the fourth learnable parameter matrix to obtain the local root cause features.

4. The method for predicting the root cause of a failure as described in claim 1, characterized in that, The hyper-relationship knowledge graph is constructed through the following steps: Candidate entities corresponding to the fault phenomenon information are determined from the preset knowledge graph, wherein the candidate entities are a set of entities distinguished according to the ontology class; Based on the text features of the fault phenomenon information and the similarity between the text features of each entity in the candidate entities corresponding to the fault phenomenon information, the non-root cause entity or the root cause entity corresponding to the fault phenomenon information is determined to obtain the case superedge of the super relation knowledge graph. Clustering is performed based on the textual features of the root cause entities in the preset knowledge graph, and the root cause entities of the same cluster are taken as the entities contained in the root cause class super edge of the super relation knowledge graph.

5. The method for predicting the root cause of a failure as described in claim 2, characterized in that, The step of extracting global case features of the case hyperedge and global root cause class features of the root cause class hyperedge using the first learnable parameter matrix and the second learnable parameter matrix, respectively, includes: By extracting the text features of the fault cases and the structural features within the case hyperedges, and combining the extracted text features and structural features with the first learnable parameter matrix, the global case features of the case hyperedges are obtained. By extracting the textual and structural features within the root cause hyperedge, and combining the extracted textual and structural features with the second learnable parameter matrix, the global root cause features of the root cause hyperedge are obtained.

6. A training method for a root cause prediction model, characterized in that, The training method for the root cause prediction model includes: Acquire training data and a hyper-relationship knowledge graph, wherein the training data includes fault cases for training, each fault case containing at least one fault phenomenon information, the hyper-relationship knowledge graph is pre-constructed based on the fault cases and a preset knowledge graph, the preset knowledge graph contains non-root cause entities and root cause entities, the hyper-relationship knowledge graph contains case hyperedges and root cause class hyperedges, the case hyperedges contain the non-root cause entity or the root cause entity corresponding to each fault phenomenon information in the fault case, and the root cause class hyperedges contain the root cause entities of the same class obtained by clustering; The fault cases and the hyperrelationship knowledge graph are input into the fault root cause prediction model to obtain the predicted value of the fault root cause. Calculate the first loss between the predicted value and the corresponding true value of the root cause of the fault; The fault root cause prediction model is trained based on the first loss.

7. The training method as described in claim 6, characterized in that, The root cause prediction model includes multiple learnable parameter matrices. The step of inputting the fault cases and the hyperrelationship knowledge graph into the fault root cause prediction model to obtain the predicted value of the fault root cause includes: The fault cases and the hyperrelationship knowledge graph are input into the fault root cause prediction model, and the fault root cause prediction model performs the following operations: Using the first learnable parameter matrix and the second learnable parameter matrix, the global case features of the case hyperedge and the global root cause class features of the root cause class hyperedge are extracted respectively. Based on the dot product of the global case features and the global root cause class features, the predicted value of the root cause class of the fault root cause is obtained. Local case features are obtained by filtering the global case features based on the third learnable parameter matrix; Based on the filtering of the global root cause class features using the fourth learnable parameter matrix, local root cause features are obtained. Based on the predicted value of the root cause class, the weights of the corresponding root causes in the local root cause features are adjusted to obtain the weight-adjusted local root cause features. The predicted value of the fault root cause is obtained based on the dot product between the local case features and the weighted local root cause features. The step of training the fault root cause prediction model based on the first loss includes: The learnable parameter matrix in the fault root cause prediction model is trained based on the first loss.

8. The training method as described in claim 7, characterized in that, The method further includes: Calculate the second loss between the predicted value and the corresponding true value of the root cause class; The step of training the fault root cause prediction model based on the first loss includes: The multiple learnable parameter matrices in the fault root cause prediction model are trained based on the first loss and the second loss.

9. The training method as described in claim 7, characterized in that, The method further includes: The third loss is calculated through contrastive learning based on the similarity between the text features of the fault cases and the text features of the root cause entities in the hyperrelationship knowledge graph. The step of training the fault root cause prediction model based on the first loss includes: The multiple learnable parameter matrices in the fault root cause prediction model are trained based on the first loss and the third loss.

10. The training method as described in claim 8, characterized in that, The method further includes: The third loss is calculated through contrastive learning based on the similarity between the text features of the fault cases and the text features of the root cause entities in the hyperrelationship knowledge graph. The step of training the plurality of learnable parameter matrices in the fault root cause prediction model based on the first loss and the second loss includes: The multiple learnable parameter matrices in the fault root cause prediction model are trained based on the first loss, the second loss, and the third loss.

11. A device for predicting the root cause of a failure, characterized in that, The fault root cause prediction device includes: The data acquisition unit is configured to acquire fault descriptions and hyperrelationship knowledge graphs, wherein the fault descriptions include at least one fault phenomenon information, the hyperrelationship knowledge graphs are pre-constructed based on fault cases and preset knowledge graphs, the fault cases include at least one fault phenomenon information, the preset knowledge graphs include non-root cause entities and root cause entities, the hyperrelationship knowledge graphs include case hyperedges and root cause class hyperedges, the case hyperedges include the non-root cause entity or the root cause entity corresponding to each fault phenomenon information in the fault case, and the root cause class hyperedges include the root cause entities of the same class obtained by clustering; The root cause prediction unit is configured to input the fault description and the hyperrelationship knowledge graph into the fault root cause prediction model to obtain the predicted value of the fault root cause.

12. A training device for a root cause prediction model, characterized in that, The training device for the root cause prediction model includes: The data acquisition unit is configured to acquire training data and a hyper-relationship knowledge graph, wherein the training data includes fault cases for training, each fault case containing at least one fault phenomenon information, the hyper-relationship knowledge graph is pre-constructed based on the fault cases and a preset knowledge graph, the preset knowledge graph contains non-root cause entities and root cause entities, the hyper-relationship knowledge graph contains case hyperedges and root cause class hyperedges, the case hyperedges contain the non-root cause entity or the root cause entity corresponding to each fault phenomenon information in the fault case, and the root cause class hyperedges contain the root cause entities of the same class obtained by clustering; The root cause prediction unit is configured to input the fault cases and the hyperrelationship knowledge graph into the fault root cause prediction model to obtain the predicted value of the fault root cause. The first loss calculation unit is configured to: calculate the first loss between the predicted value and the corresponding true value of the root cause of the fault; The model training unit is configured to train the fault root cause prediction model based on the first loss.

13. An electronic device, characterized in that, include: At least one processor; At least one memory that stores computer-executable instructions. Wherein, when the computer-executable instructions are executed by the at least one processor, the at least one processor causes the at least one processor to execute the fault root cause prediction method as described in any one of claims 1-5 and the fault root cause prediction model training method as described in any one of claims 6-10.

14. A computer-readable storage medium for storing instructions, characterized in that, When the instruction is executed by at least one processor, it causes the at least one processor to perform the root cause prediction method as described in any one of claims 1-5 and the root cause prediction model training method as described in any one of claims 6-10.

15. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the method for predicting the root cause of a fault as described in any one of claims 1-5 and the method for training the root cause prediction model as described in any one of claims 6-10.

Images