An industrial system fault diagnosis method based on a knowledge graph

Abstract

Embodiments of the present disclosure relate to a method for industrial system fault diagnosis based on a knowledge graph. The method comprises: constructing a knowledge graph database and a plurality of fault types based on component relationship data, operation data and fault data of an industrial system, performing data completion and alignment on the knowledge graph database to form a fault information knowledge graph; constructing training samples and test samples according to the fault information knowledge graph and the fault types; predicting each fault type and constructing a corresponding first fault diagnosis model; inputting the training samples into the first fault diagnosis model for training, optimizing the trained first fault diagnosis model according to the test samples to obtain a second fault diagnosis model; and constructing a global fault diagnosis model in combination with all the second fault diagnosis models. The embodiments of the present disclosure can mine the correlation between the component operating conditions and the faults of a specific industrial system, and guide system experts to infer the faults of the industrial system.

Description

Technical Field

[0001] This disclosure relates to the field of knowledge mining technology, and in particular to a method for diagnosing industrial system faults based on knowledge graphs. Background Technology

[0002] With the development of artificial intelligence technology and the advancement of automation, industrial systems have been widely applied in various fields. Industrial systems consist of a large number of interconnected components and subsystems, and their behavior exhibits high nonlinearity, dynamism, and uncertainty. However, due to the complexity of these systems, traditional fault diagnosis methods often struggle to accurately determine the root cause and location of faults when they occur, leading to cumbersome and time-consuming troubleshooting processes and a high risk of misdiagnosis.

[0003] Fault diagnosis in industrial systems faces several challenges. First, industrial systems typically consist of numerous components and subsystems with complex dependencies and interactions. This interrelationship complexity leads to diverse and complex fault propagation paths, increasing the difficulty of fault diagnosis. Second, the operating state of industrial systems is often uncertain and can be affected by various factors, such as environmental changes and component aging. This uncertainty necessitates considering numerous uncertainties in the fault diagnosis process. Furthermore, industrial systems are usually highly dynamic, with their operating states constantly changing over time; therefore, fault diagnosis methods need to be real-time and adaptive. Summary of the Invention

[0004] To avoid the shortcomings of existing technologies, this application provides a knowledge graph-based industrial system fault diagnosis method to solve the problem that existing technologies have difficulties in accurately determining the root cause and location of faults, resulting in cumbersome and time-consuming troubleshooting processes and a high risk of misdiagnosis.

[0005] According to embodiments of this disclosure, a fault diagnosis method for industrial systems based on knowledge graphs is provided, the method comprising:

[0006] A knowledge graph database is constructed based on component relationship data, operation data, and fault data of industrial systems. Multiple fault types are obtained, and the knowledge graph database is supplemented and aligned to form a fault information knowledge graph.

[0007] Training and test samples are constructed based on the fault information knowledge graph and the fault types.

[0008] For each of the aforementioned fault types, a prediction is made, and a corresponding first fault diagnosis model is constructed based on the prediction results.

[0009] The training samples are input into the first fault diagnosis model for training, and the trained first fault diagnosis model is optimized based on the test samples to obtain the second fault diagnosis model.

[0010] A global fault diagnosis model is constructed by combining all the second fault diagnosis models.

[0011] The step of constructing a knowledge graph database based on component relationship data, operational data, and fault data of an industrial system, obtaining multiple fault types, and performing data completion and alignment on the knowledge graph database to form a fault information knowledge graph includes:

[0012] Based on the system model, the system entity, the models of different system components, the manufacturers of different components, the place of origin of different manufacturers, and the relationship between components, a system firmware relationship diagram is generated.

[0013] Based on the system's historical fault records and the location of the faults, and combined with the system firmware relationship diagram, an industrial system fault diagram database is generated.

[0014] Based on the similarity between historical fault records of different systems in the industrial system fault graph database, various fault types can be obtained;

[0015] The fault types are added as nodes to the industrial system fault graph database to generate an industrial system fault knowledge graph.

[0016] The data completion and alignment are performed on the industrial system fault knowledge graph to generate the industrial system fault information knowledge graph.

[0017] The step of adding the fault type as a node to the industrial system fault graph database to generate an industrial system fault knowledge graph includes:

[0018] Based on the similarity between historical fault records of different systems in the industrial system fault graph database, the faults are clustered into several groups.

[0019] If the similarity within a group is greater than a preset threshold, it is considered to be the same type of fault.

[0020] The step of constructing training and testing samples based on the fault information knowledge graph and the fault type includes:

[0021] The normal operation records of the industrial system are divided into preset time periods to generate a normal dataset;

[0022] The number of faults corresponding to each of the aforementioned fault types is counted, and the normal dataset is randomly sampled based on the number of faults for each of the aforementioned fault types.

[0023] The nodes and attributes corresponding to each fault type are separated from the industrial system fault information knowledge graph as a subgraph to form a fault node set, thus constituting a fault sample set.

[0024] The attributes in the fault sample set are replaced with the attribute values ​​of the non-fault samples obtained by sampling the normal dataset to form a non-fault sample set.

[0025] The fault sample set and the non-fault sample set are divided according to a preset ratio to obtain the training samples and the test samples.

[0026] The operation records of the industrial system are divided into segments of 0.1 seconds.

[0027] The ratio of the training samples to the test samples is 3:1.

[0028] The step of predicting for each of the aforementioned fault types and constructing a corresponding first fault diagnosis model based on the prediction results includes:

[0029] The probability of occurrence of each of the aforementioned fault types is predicted, and the first fault diagnosis model is modeled based on the prediction results and the C4.5 decision tree algorithm.

[0030] The step of inputting the training samples into the first fault diagnosis model for training, and optimizing the trained first fault diagnosis model based on the test samples to obtain the second fault diagnosis model includes:

[0031] The training samples are input into the first fault diagnosis model for training.

[0032] Based on the test samples, the fault diagnosis decision tree of the first fault diagnosis model is pruned to obtain the second fault diagnosis model.

[0033] The method also includes:

[0034] The data to be diagnosed is matched with the data of the training samples corresponding to the second fault diagnosis model to be judged. The data is then filled and aligned. For missing attribute values, the discrete attribute preset value and the continuous attribute preset value are calculated and filled respectively. For extra attributes, the corresponding attribute is deleted.

[0035] The data to be diagnosed, after being completed and aligned, is input into the global fault diagnosis model to obtain the probability that the data belongs to different fault types, and the fault type of the data to be diagnosed is determined based on the probability.

[0036] The technical solutions provided by the embodiments of this disclosure may include the following beneficial effects:

[0037] In the embodiments of this disclosure, the aforementioned knowledge graph-based industrial system fault diagnosis method, on the one hand, constructs an industrial system fault knowledge graph connecting component relationships and historical fault data, providing data support for fault diagnosis in the industrial system. Then, based on the industrial system fault knowledge graph, it learns fault type classifications and fault diagnosis models for different fault types, and applies these models to the diagnosis of fault data. On the other hand, by constructing a knowledge graph of fault data and learning fault diagnosis models, it avoids the instability of human experience and the inability to identify deeper feature relationships inherent in traditional expert-based fault inference methods, which can lead to unstable fault diagnosis results and affect system operation. Attached Figure Description

[0038] 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. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0039] Figure 1 The diagram illustrates the steps of a knowledge graph-based industrial system fault diagnosis method according to an exemplary embodiment of this disclosure.

[0040] Figure 2 A flowchart illustrating a knowledge graph-based industrial system fault diagnosis method in an exemplary embodiment of this disclosure is shown. Detailed Implementation

[0041] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that this disclosure will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0042] Furthermore, the accompanying drawings are merely illustrative diagrams of embodiments of this disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities.

[0043] This example implementation provides a knowledge graph-based method for fault diagnosis of industrial systems. (Reference) Figure 1 As shown, the knowledge graph-based industrial system fault diagnosis method may include steps S101 to S105.

[0044] Step S101: Construct a knowledge graph database based on component relationship data, operation data, and fault data of the industrial system, and obtain multiple fault types. Perform data completion and alignment on the knowledge graph database to form a fault information knowledge graph.

[0045] Step S102: Construct training samples and test samples based on the fault information knowledge graph and the fault type;

[0046] Step S103: For each of the aforementioned fault types, make a prediction and construct a corresponding first fault diagnosis model based on the prediction results;

[0047] Step S104: Input the training samples into the first fault diagnosis model for training, and optimize the trained first fault diagnosis model according to the test samples to obtain the second fault diagnosis model.

[0048] Step S105: Construct a global fault diagnosis model by combining all the second fault diagnosis models.

[0049] The aforementioned knowledge graph-based industrial system fault diagnosis method addresses two key issues. First, it uncovers the correlation between the operational status of components and fault occurrences within a specific industrial system, guiding system experts in inferring faults. Second, it solves the problem of constructing a fault information knowledge graph connecting industrial system components and fault cases, including entity extraction, data alignment, and completion of structured data. Furthermore, based on the constructed fault information knowledge graph, it designs related secondary fault diagnosis models and a global fault diagnosis model to achieve fault category determination of fault data.

[0050] Below, we will refer to Figures 1 to 2 The steps of the knowledge graph-based industrial system fault diagnosis method described in this example embodiment will be explained in more detail.

[0051] In step S101, a graph database is constructed based on existing industrial system component relationship data, system operation data, and system fault data. Through data completion and alignment, a fault information knowledge graph of the industrial system is formed, specifically as follows:

[0052] Step S1011: Take the system model, system entity, model of different system components, manufacturer of different components, place of origin of different manufacturers, and component as nodes, and take the relationships between system model and system entity, system model and model of different system components, model of different components and manufacturer, different manufacturers and place of origin, system entity and different system components, and model of component as components, and generate system firmware relationship diagram G3.

[0053] Step S1012: The fault history record number and the location of the fault are used as nodes. The occurrence time of different faults and the records before and after the faults of multiple sensors are used as attributes of the fault history record number node. The relevant components of the fault corresponding to the fault history record number are used as relationships and merged with the system firmware relationship graph G3 obtained in step S1011 to generate the industrial system fault graph database G2.

[0054] Step S1013: By calculating the similarity between historical fault records of different systems in the industrial system fault graph database generated in step S1012, the faults are clustered into N groups. Groups with high similarity within a group are considered to be the same type of fault, while groups with low similarity are considered to be different types of faults.

[0055] Step S1014: Add the N fault types generated in step S1013 as nodes to the industrial system fault graph database generated in step S1012 to generate an industrial system fault knowledge graph G1 with preliminary inference information.

[0056] Step S1015: Process N types of faults separately. Fault types of the same type have a uniform number and category of attributes. Find missing attributes, set different default values ​​for different attribute data types, and use the default values ​​of that attribute type to complete the corresponding faults. After completing and aligning the knowledge graph G1 generated in step S1014, a complete industrial system fault information knowledge graph G is obtained.

[0057] In steps S102 to S105, based on the industrial system fault information knowledge graph G and the historical operation information of the industrial system generated in step S1015, fault training samples and test samples are constructed, and a fault diagnosis model based on C4.5 decision tree is designed, specifically as follows:

[0058] The data from normal system operation is divided into 0.1-second intervals to generate a normal dataset ND. The number N of faults corresponding to the N fault types generated in step S1013 is counted. i Where i corresponds to the i-th fault type, the normal dataset ND is randomly sampled and compared with each fault type N. i Equal sample size D i .

[0059] Each fault's corresponding node and attribute are separated from the industrial system fault information knowledge graph G generated in step S1015 into a fault node set, which constitutes the fault sample set. Simultaneously, the attributes in the fault sample set are replaced with N obtained by sampling the normal dataset ND. i The non-faulty sample attribute values ​​constitute the non-faulty sample set NG. i G i and NG i The samples were divided into training samples T according to a sampling ratio of 3:1. i and test sample V i ;

[0060] The training samples T i A fault diagnosis model is built based on the C4.5 decision tree algorithm, and the model is also based on test samples V. i Pruning is performed on the fault diagnosis decision tree to ensure its generalization performance. For N fault categories, the corresponding dataset and the fault diagnosis decision tree for the i-th category are obtained respectively. The branches of the N fault diagnosis decision trees are combined into a global fault diagnosis model.

[0061] Obtain the fault data to be diagnosed, perform feature completion and alignment on the input fault according to the data alignment strategy on which N types of faults depend, input the global fault diagnosis model, and output the probability of the fault data belonging to each of the N types of faults.

[0062] The probabilities of the N types of faults obtained represent the likelihood that the fault corresponding to the current fault data belongs to different fault categories. All fault category probabilities are evaluated; those with a probability greater than 0.9 are identified as fault categories by the fault diagnosis model. After obtaining the fault categories from the fault data, the corresponding fault troubleshooting procedures are applied to each category for subsequent fault troubleshooting work.

[0063] This application provides a knowledge graph-based method for fault diagnosis in industrial systems. It aims to uncover the correlation between the operational status of components and the occurrence of faults in specific industrial systems, guiding system experts to infer faults. The main challenges are: firstly, constructing a knowledge graph connecting industrial system components and fault cases, including entity extraction, data alignment, and completion of structured data; secondly, designing a relevant fault diagnosis model based on the constructed knowledge graph to achieve fault category determination of fault data.

[0064] The present application will be further described below with reference to the accompanying drawings and embodiments, which include, but are not limited to, the following embodiments.

[0065] like Figure 2 As shown, this application provides a knowledge graph-based method for fault diagnosis of industrial systems, the basic implementation process of which is as follows:

[0066] 1. Construction of a knowledge graph for industrial system fault information

[0067] Based on existing relational databases of industrial system components, this method transforms the potential relationships between tables into a graph of nodes and relationships by predefining the relationships between nodes. By completing and aligning fault-related key data, the system firmware relationship graph is integrated with the fault information of the system, forming a complete knowledge graph of industrial system fault information.

[0068] (1) Construction of system component relationship diagram G3

[0069] The industrial system primarily consists of three tables: Entity Table T1, Component Table T2, and Fault Table T3. The Entity Table has attributes including Entity Number (primary key), System Model, Component Number, and Component Model. The Component Table has attributes including Component Model (primary key), Origin, and Manufacturer. The Fault Table has attributes including Fault Number (primary key), Entity Number (primary key), Component Number (primary key), Work Location, and n attributes. A single or multiple primary key attributes uniquely determine the uniqueness of each record. Attributes with the same name across different tables act as foreign key constraints, representing the relationships between the tables.

[0070] First, a system component relationship graph G3 is constructed between entity table T1 and component table T2. Component model numbers serve as constraints for the relationships between the two tables. This graph is represented using RDF (Resource Description Framework). Entities, component numbers, component models, places of origin, and manufacturers are represented as nodes. Relationships between nodes are established using records from each row of tables T1 and T2, reducing redundancy in the system's component representation relationships. The specific implementation is shown below:

[0071] Entity table T1

[0072]

[0073] Entity table T2

[0074]

[0075] RDF (Relational Principle) represents (Triple Relation):

[0076] R1 = {(Entity: S001) - [Contains] -> (Component: P0)}

[0077] R2 = {(Entity: S001) - [Contains] -> (Component: P1)}

[0078] R3 = {(Entity: S002) - [Contains] -> (Component: P2)}

[0079] R4 = {(Component Model: L2) - [Origin] -> (Place of Origin: Z1)}

[0080] R5 = {(Component Model: L3) - [Origin] -> (Place of Origin: Z2)}

[0081] R6 = {(Component Model: L5) - [Origin] -> (Place of Origin: Z2)}

[0082] R7 = {(Component Model: L2) - [Production] -> (Origin: Q)}

[0083] R7 = {(Component Model: L3) - [Production] -> (Origin: W)}

[0084] R8 = {(Component Model: L5) - [Production] -> (Origin: W)}

[0085] (2) Construction of G2 Fault Graph Database for Industrial Systems

[0086] Similar to the strategy of merging tables T1 and T2 in (1), based on RDF (Resource Description Framework), each record represented by fault table T3 is merged into G3. The nodes extracted from table T3 are fault number and work site, and the attributes associated with tables T1 and T2 are entity number and part number, respectively. The same fault number may be associated with multiple part numbers due to the specific nature of the fault. Unlike the triplet relationship in (1), the fault node has n attributes, some of which are missing. The specific representation is as follows:

[0087] R9 = {(Fault: G5{Attribute 1: x1; Attribute 2: x2; ...}) - [Association] -> (Part Number: P0)}

[0088] The component relationship diagram generated in (2) is merged with the fault node triplet to form the industrial system fault diagram database G2.

[0089] (3) Construction of Knowledge Graph G for Fault Information in Industrial Systems

[0090] Because the fault nodes in the industrial system fault graph database G2 obtained in (2) are independent of each other and lack correlation, this poses a significant obstacle to fault diagnosis. This summary proposes an index to evaluate the correlation between faults by fusing topological similarity estimation and node attribute similarity estimation. The specific details are as follows:

[0091] Topological similarity calculation:

[0092]

[0093] Among them, S i,j R represents the topological similarity between the i-th and j-th faulty nodes; i and R j Let represent the number of adjacent nodes of the i-th and j-th faulty nodes, respectively. The closer the similarity is to 1, the higher the topological similarity between the two faults.

[0094] Calculation of similarity of attributes of fault nodes:

[0095]

[0096] Among them, A i,j This represents the attribute similarity between the i-th and j-th fault nodes; k represents the number of attributes possessed by each fault node. and Let represent the values ​​of the k-th attribute of the i-th and j-th fault nodes, respectively; Φ represents the function for calculating similarity based on discrete and continuous attribute values, as shown in the following formula:

[0097]

[0098] Obtain the topological similarity S among all faulty nodes. i,j Similarity to attributes A i,j Combining the product of the two metrics yields the consistent similarity C. i,j The specific calculation method is as follows:

[0099] C i,j =A i,j ×S i,j

[0100] The consistency similarity between all pairs of nodes is obtained, and a fault association subgraph G consisting of all faulty nodes is obtained based on the consistency similarity. s A threshold t = 0.75 is set. If the consistency similarity between faulty nodes is greater than the threshold t, the relationship between the two points is retained; otherwise, the relationship between the two points is removed. Finally, G... sDivide into N unrelated subgraphs. All faults within these N subgraphs belong to the same fault category; the subgraphs belong to different fault categories. Generate N fault category nodes, add these N fault category nodes to the industrial system fault graph database G2 generated in (2), and connect all fault nodes with their corresponding fault category nodes to form an industrial system fault knowledge graph G1 with preliminary inference information.

[0101] Because identical fault nodes may exhibit misalignment in their attributes due to differences in occurrence time and measurement conditions, misalignment can occur for the same type of fault. To address this limitation in data sampling, attribute completion and alignment are performed on both discrete and continuous samples: For discrete samples, the attribute values ​​of all corresponding attributes for the same type of fault nodes are statistically analyzed, and the most frequently occurring attribute category is used as the preset attribute value; for continuous samples, the attribute values ​​of all corresponding attributes for the same type of fault nodes are averaged, and this average is used as the preset attribute value. Then, all missing fault attributes are completed using the statistically derived preset attribute values, resulting in a complete industrial system fault information knowledge graph G.

[0102] 2. Design of Fault Diagnosis Model for Industrial Systems

[0103] Historical operational data from sampled industrial systems are used to construct a dataset suitable for training a fault diagnosis model through enhancement and balancing of test and training examples. After training on the training set (i.e., training samples) and validating on the test set (i.e., test samples), a fault diagnosis model with high generalization performance is obtained. The specific steps are as follows:

[0104] (1) ND partitioning of normal dataset

[0105] The operational records of industrial systems are stored with a minimum granularity of 0.1 seconds. This method primarily addresses fault diagnosis of system operation at a specific moment. Therefore, the operational records of the industrial system are divided into all historical operational records using 0.1-second intervals as the dividing points. When a system malfunctions, an anomaly occurs, and the historical operational records are specially marked. Therefore, by default, all acquired operational records are considered anomaly-free data. After dividing all operational records, a normalized dataset ND is obtained. Each instance of ND contains all attributes related to system operation.

[0106] (2) Division of training examples (i.e., training samples) and test examples (i.e., test samples)

[0107] Iterate through the N fault categories obtained in step 1, and for the i-th fault category, the number of faults is N. i Sample from the normal dataset ND and N i Equal D iN samples are used as the candidate fault-free dataset for the i-th fault category. To ensure consistency between fault and non-fault data across different attributes, the preset values ​​of each attribute calculated in step 1 are used as preset values ​​for the non-fault dataset to complete and align it, while attributes irrelevant to the i-th fault category are removed. In this way, N fault samples are obtained for each fault category. i Dataset ND with the same number of non-faulty samples i .

[0108] Machine learning methods require both training and testing sets to train and validate the model, respectively, and these sets need to maintain a certain degree of consistency in distribution. The obtained datasets of faulty and non-faulty samples are evenly divided in a 3:1 ratio, with the faulty and non-faulty samples comprising 3 / 4 of the dataset as the training set T. i The remaining 1 / 4 constitutes the test set V. i To test the generalization performance of the fault diagnosis model.

[0109] (3) Generation of Fault Diagnosis Decision Tree Model

[0110] This method first performs independent fault category prediction for each fault category, and then combines all fault category diagnostic models to generate a global fault diagnosis model (i.e., a global fault diagnosis model), satisfying the system's scalability. For the prediction of the fault occurrence probability Pi for a single fault category, a C4.5 decision tree model is used for training and prediction. The decision tree generation process is shown below:

[0111] Input: Training samples D = {(x1,y1),(x2,y2),...}, A = {a1,a2,...}. Here, D represents the training sample x and fault representation y for a certain fault category, where y = 1 indicates the sample is a faulty sample, and y = 0 indicates the sample is a non-faulty sample; A represents the unclassified attribute set of all samples.

[0112] a. Generate a node

[0113] b. If all nodes in D have the same y-value, then the node is identified as belonging to that same category; End;

[0114] c. If an attribute in A is empty, or if all elements in D have the same value for every attribute, then set the node to the category with the most elements in D; end.

[0115] d. Calculate the remaining attributes based on the C4.5 algorithm to obtain the attribute a* with the largest information gain;

[0116] e. Iterate through each attribute value in a*. If the number of sample divisions for a certain attribute value is empty, then set that node to the class with the most frequent division in D; end. Otherwise, input D and A\{a*} as new training samples and recalculate from step a.

[0117] f. Once all branches have finished running, the decision tree is complete.

[0118] Decision trees based on training sets often carry the risk of overfitting and need to be validated using test sets to assess their generalization performance. A validation performance threshold of 85% accuracy is set. First, a global test is performed on each example in the test set to determine whether the judgment is correct or incorrect. The judgment results for all test examples of that category are then statistically analyzed. If the accuracy is higher than 85%, the validation passes. If it is lower than 85%, post-pruning is performed, as follows:

[0119] Traverse each leaf node of the decision tree. If a leaf node is removed, all nodes of its parent node become leaf nodes, and its class is set to the class with the highest percentage among all samples in its branch. Recalculate the validation accuracy for that class. If there is an improvement, prune the tree (i.e., perform the above operation). If there is no improvement, check other leaf nodes. When the total height of the decision tree is less than or equal to 3, delete the current decision tree and recalculate the training samples T of the dataset. i and test case V i The data is divided and then regenerated through a decision tree generation process until the accuracy of the validation set meets the requirements.

[0120] 3. Fault Category Inference for Industrial Systems

[0121] During the operation of industrial systems, the abnormal phenomena they exhibit can be influenced by various factors, resulting in complex fault categories that cannot be easily determined by individual attributes. Therefore, fault data is input into all fault diagnosis decision trees to obtain the probabilities of different fault categories, thereby determining the fault category of the industrial system. Obtaining abnormal operation data involves the following two steps:

[0122] (1) Completion and alignment of data to be diagnosed

[0123] Different fault categories may have significant differences in their anomaly-related attributes. After obtaining the anomaly data, the attributes of the anomaly data are first completed and aligned with the training samples corresponding to the decision tree for the fault category to be judged. For missing attribute values, they are filled based on the preset values ​​of discrete and continuous attributes calculated in step 1; for extra attributes, the corresponding attributes are deleted.

[0124] (2) Fault Category Probability Prediction

[0125] The completed and aligned data obtained in (1) are input into decision trees for different fault types, and the probability of the input data belonging to different fault categories is finally given. The output probability is filtered according to a threshold of 0.90, and all fault categories with a probability greater than 0.90 are the fault categories of the input data.

[0126] The aforementioned knowledge graph-based industrial system fault diagnosis method addresses two key issues. First, it uncovers the correlation between the operational status of components and fault occurrences within a specific industrial system, guiding system experts in inferring faults. Second, it solves the problem of constructing a fault information knowledge graph connecting industrial system components and fault cases, including entity extraction, data alignment, and completion of structured data. Furthermore, based on the constructed fault information knowledge graph, it designs related secondary fault diagnosis models and a global fault diagnosis model to achieve fault category determination of fault data.

[0127] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of this disclosure, "a plurality of" means two or more, unless otherwise explicitly specified.

[0128] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0129] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the appended claims.

Claims

1. A method for fault diagnosis of industrial systems based on knowledge graphs, characterized in that, The method includes: Step S101: Construct a knowledge graph database based on component relationship data, operational data, and fault data of the industrial system, and obtain various fault types. Perform data completion and alignment on the knowledge graph database to form a fault information knowledge graph; specifically including: Based on the system model, the system entity, the models of different system components, the manufacturers of different components, the place of origin of different manufacturers, and the relationship between components, a system firmware relationship diagram is generated. Based on the system's historical fault records and the location of the faults, and combined with the system firmware relationship diagram, an industrial system fault diagram database is generated. Based on the similarity between historical fault records of different systems in the industrial system fault graph database, various fault types can be obtained; The fault types are added as nodes to the industrial system fault graph database to generate an industrial system fault knowledge graph. The faults are clustered into several groups based on the similarity between nodes of historical fault records in different systems in the industrial system fault graph database. If the similarity within a group is greater than a preset threshold, it is considered to be the same fault type. The data completion and alignment are performed on the industrial system fault knowledge graph to generate the industrial system fault information knowledge graph. Step S102: Construct training samples and test samples based on the fault information knowledge graph and the fault type; specifically including: The normal operation records of the industrial system are divided into preset time periods to generate a normal dataset; The number of faults corresponding to each of the aforementioned fault types is counted, and the normal dataset is randomly sampled based on the number of faults for each of the aforementioned fault types. The nodes and attributes corresponding to each fault type are separated from the industrial system fault information knowledge graph as a subgraph to form a fault node set, thus constituting a fault sample set. The attributes in the fault sample set are replaced with the attribute values ​​of the non-fault samples obtained by sampling the normal dataset to form a non-fault sample set. The fault sample set and the non-fault sample set are divided according to a preset ratio to obtain the training samples and the test samples; Step S103: For each of the aforementioned fault types, make a prediction and construct a corresponding first fault diagnosis model based on the prediction results; Step S104: Input the training samples into the first fault diagnosis model for training, and optimize the trained first fault diagnosis model according to the test samples to obtain the second fault diagnosis model. Step S105: Construct a global fault diagnosis model by combining all the second fault diagnosis models.

2. The industrial system fault diagnosis method based on knowledge graphs according to claim 1, characterized in that, The operation records of the industrial system are divided into segments of 0.1 seconds.

3. The industrial system fault diagnosis method based on knowledge graphs according to claim 2, characterized in that, The ratio of the training samples to the test samples is 3:

1.

4. The industrial system fault diagnosis method based on knowledge graphs according to claim 3, characterized in that, The step of predicting for each of the aforementioned fault types and constructing a corresponding first fault diagnosis model based on the prediction results includes: The probability of occurrence of each of the aforementioned fault types is predicted, and the first fault diagnosis model is modeled based on the prediction results and the C4.5 decision tree algorithm.

5. The industrial system fault diagnosis method based on knowledge graphs according to claim 4, characterized in that, The step of inputting the training samples into the first fault diagnosis model for training, and optimizing the trained first fault diagnosis model based on the test samples to obtain the second fault diagnosis model includes: The training samples are input into the first fault diagnosis model for training. Based on the test samples, the fault diagnosis decision tree of the first fault diagnosis model is pruned to obtain the second fault diagnosis model.

6. The industrial system fault diagnosis method based on knowledge graphs according to claim 5, characterized in that, The method also includes: The data to be diagnosed is matched with the data of the training samples corresponding to the second fault diagnosis model to be judged. The data is then filled and aligned. For missing attribute values, the discrete attribute preset value and the continuous attribute preset value are calculated and filled respectively. For extra attributes, the corresponding attribute is deleted. The data to be diagnosed, after being completed and aligned, is input into the global fault diagnosis model to obtain the probability that the data belongs to different fault types, and the fault type of the data to be diagnosed is determined based on the probability.

Images