Microwave filter diagnosis and repair inversion method based on knowledge graph

Abstract

The application discloses a microwave filter diagnosis and repair inversion method based on a knowledge graph, and the steps are as follows: obtaining full-period unstructured text data of a microwave filter, and constructing an ontology of a microwave filter knowledge graph; preprocessing the unstructured text data to generate a training set and a test set; training an entity extraction model based on a bidirectional long short-term memory network, associating the extracted entities according to the ontology, obtaining knowledge triples, and constructing a microwave filter knowledge graph; fusing the microwave filter knowledge graph with a diagnosis and repair inversion model; and based on the microwave filter knowledge graph, performing diagnosis and repair inversion to realize fault diagnosis, debugging guidance and design. The application improves the utilization rate of accumulated data of the microwave filter, enhances the cooperation of microwave filter design and manufacturing, and improves the design and manufacturing iteration efficiency.

Description

Technical Field

[0001] This invention belongs to the field of physical technology, and further relates to a knowledge graph-based method for microwave filter diagnosis and repair inversion within the field of data processing technology. This invention can construct a microwave filter knowledge graph based on the full-cycle unstructured data of the microwave filter, and realize the association of full-cycle data and fault and repair inversion based on the knowledge graph. Background Technology

[0002] Knowledge graphs are structured semantic knowledge bases that describe concepts and their relationships in the physical world in symbolic form. They enable the association and explicit representation of massive amounts of data. Microwave filters, as frequency-selective devices, directly impact the communication quality of the entire system due to their electrical performance; design and manufacturing are critical steps. Microwave filters generate a large amount of unstructured text data during the design and manufacturing stages, which is stored in the enterprise's storage system. Currently, microwave filter design and manufacturing are disconnected, with no data correlation throughout the entire lifecycle. Data from the design and manufacturing stages cannot be effectively utilized, and the manufacturing stage cannot push effective information to the design stage, resulting in long feedback cycles from manufacturing to design and low design-manufacturing iteration efficiency. Effectively associating, visualizing, and implementing diagnostic and repair inversion of this data is of great significance for shortening the design and manufacturing cycle and enhancing design-manufacturing collaboration. Existing knowledge graph construction methods and inversion techniques cannot construct a microwave filter knowledge graph based on the unstructured text data throughout the entire microwave filter lifecycle, nor can they achieve diagnostic and repair inversion based on the microwave filter knowledge graph.

[0003] Suzhou Xinmai Intelligent Technology Co., Ltd. disclosed a method for transforming manual experience in the cavity filter debugging process into a knowledge graph to assist in cavity filter debugging in its patent application "An Intelligent Auxiliary Debugging and Device for Cavity Filters" (Application No. 202110609307.1, Publication No. CN 113255222 A). The method's implementation steps are as follows: First, based on the cavity filter's design specifications, obtain the ideal dissipation parameters of the cavity filter using a vector network analyzer. Second, perform a screw-in operation on the cavity filter and record the screw-in depth. Third, obtain the real-time dissipation parameters of the cavity filter after screw-in. Fourth, adjust the screw-in depth based on the difference between the real-time dissipation parameters and the ideal dissipation parameters after screw-in, and obtain and record the updated screw-in depth. Fifth, perform knowledge transformation on the screw-in depth transformation operation during the debugging process. Sixth, extract knowledge of the cavity filter's own characteristics and fuse it with the screw-in depth relationship to form a cavity filter debugging experience knowledge graph. The drawback of this method is that it only constructs a knowledge graph that correlates the dissipation parameters, characteristics, and adjustment parameters of the cavity filter. It is suitable for the adjustment step in the design and manufacturing process of the cavity filter. This knowledge graph construction method is inefficient and cannot correlate the data throughout the entire lifecycle of microwave filter design and manufacturing. It is difficult to apply to scenarios with a large amount of unstructured data. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of the prior art by proposing a knowledge graph-based method for the diagnosis and repair of microwave filters. This method solves the problems of data disconnection and information asymmetry in the design and manufacturing cycle of microwave filters, which prevents the diagnosis and repair of microwave filters and makes it difficult to apply existing technologies to the field of microwave filters with well-defined professional knowledge.

[0005] The present invention discloses a microwave filter diagnosis and repair inversion method based on a knowledge graph. The steps are as follows: collecting unstructured text data of the entire microwave filter lifecycle to construct an ontology of the knowledge graph; preprocessing the unstructured text data to generate training and testing sets; training an entity extraction model based on a bidirectional long short-term memory network to extract the unstructured text data; associating the extracted entities according to the ontology to obtain knowledge triples and constructing a microwave filter knowledge graph. The microwave filter knowledge graph is integrated with the diagnosis and repair inversion model, and diagnosis and repair inversion are performed based on the microwave filter knowledge graph to achieve fault diagnosis, guidance for debugging and design.

[0006] To achieve the above objectives, the steps of the present invention include the following:

[0007] Step 1: Collect unstructured text data generated throughout the entire cycle of the microwave filter;

[0008] Step 2, construct the ontology of the microwave filter knowledge graph:

[0009] Based on the unstructured text data generated throughout the entire lifecycle of the microwave filter, which includes microwave filter design documents, manufacturing process records, and diagnostic and maintenance records, 28 entity types and 8 relation types are defined respectively. The 28 entity types and 8 relation types are then constructed into 28 "head entity-relationship-tail entity" knowledge triples. All entity types, relation types, and knowledge triples are combined to form the ontology of the microwave filter knowledge graph.

[0010] Step 3: Generate training and test sets:

[0011] Step 3.1: Preprocess the unstructured text data;

[0012] Step 3.2: Randomly select 25% of the text data from the unstructured text data and label the entity type of each selected text data.

[0013] Step 3.3: Generate training and testing sets from the labeled text data in an 8:2 ratio;

[0014] Step 4: Obtain a qualified entity extraction model based on a bidirectional long short-term memory network.

[0015] Step 4.1: Input the training set into the entity extraction model based on the bidirectional long short-term memory network, and use the stochastic gradient descent method to iteratively update the parameters in the entity extraction model until the loss function converges, thus obtaining the trained entity extraction model.

[0016] Step 4.2: Input the test set into the trained entity extraction model, output the entity type label predicted for each text data in the test set, calculate the accuracy of the entity type labels of all text data output by the entity extraction model, and obtain a qualified entity extraction model when the extraction accuracy of the extraction model reaches 80%.

[0017] Step 5, construct a microwave filter knowledge graph:

[0018] Step 5.1: Input all text data in the unstructured text data into a qualified entity extraction model, and output the entity of each text data and its corresponding entity type.

[0019] Step 5.2: Associate the entities in each text data according to their entity types and the knowledge triples in the ontology to obtain the knowledge triples of the microwave filter knowledge graph for that text data.

[0020] Step 5.3: Store all knowledge triples in the Neo4j graph database. Using the Neo4j graph database generation instructions, treat each knowledge triple entity as a corresponding node and the relation as an edge connecting the head entity node and the tail entity node to obtain the microwave filter knowledge graph.

[0021] Step 6: Utilize the microwave filter knowledge graph to fuse the diagnostic and repair inversion model:

[0022] Step 6.1: Extract all entities with entity types of scattering parameters and component process dimensions from the microwave filter knowledge graph; combine a set of associated scattering parameters and component process dimensions into a set of input parameters for the diagnostic and repair inversion model;

[0023] Step 6.2: Select a set of previously unselected diagnostic and repair inversion input parameters; convert the scattering parameters in the selected diagnostic and repair inversion input parameters into admittance parameters, substitute them into the vector fitting algorithm, and obtain the coupling matrix including coupling coefficients and mutual coupling coefficients according to the coupling matrix synthesis theory; substitute the component process dimensions in the input parameters and the obtained coupling matrix into the space mapping algorithm to obtain the component adjustment amount that meets the microwave filter performance index.

[0024] Step 6.3: Associate the inverted coupling matrix, self-coupling coefficient, mutual coupling coefficient, and component adjustment amount with the scattering parameter and component process size in the same set of input parameters; obtain 5 inversion knowledge triples: "scattering parameter-inversion-coupling matrix", "coupling matrix-including self-coupling coefficient", "coupling matrix-including mutual coupling coefficient", "coupling matrix-inversion-component adjustment amount", and "component process size-inversion-component adjustment amount". Store the 5 inversion knowledge triples in the Neo4j graph database to obtain the microwave filter knowledge graph after updating the inversion knowledge triples;

[0025] Step 6.4: Determine whether all input parameters for diagnosis and repair inversion have been selected. If yes, proceed to step 7; otherwise, proceed to step 6.2.

[0026] Step 7: Perform diagnosis and repair inversion based on the microwave filter knowledge graph:

[0027] Step 7.1: When a microwave filter fails, the same method as in Steps 3.1, 5.1 and 5.2 is used to extract entities from the text data of the failed microwave filter, obtain the knowledge triplet of the failed microwave filter, store the knowledge triplet of the failed microwave filter in the Neo4j graph database, and obtain the microwave filter knowledge graph after updating the knowledge triplet of the failed microwave filter.

[0028] Step 7.2: Using a similarity algorithm, match the microwave filter name and scattering parameters that are similar to those of the faulty microwave filter in the updated microwave filter knowledge graph.

[0029] Step 7.3: Determine whether the similarity between the microwave filter name in the knowledge graph and the microwave filter name of the faulty microwave filter, and the similarity between the scattering parameters in the microwave filter knowledge graph and the scattering parameters of the faulty microwave filter, meet the similarity conditions. If yes, proceed to step 7.4; otherwise, proceed to step 7.5.

[0030] Step 7.4: The debugging and design personnel search for solutions corresponding to the microwave filter name and scattering parameters in the microwave filter knowledge graph; and repair the faulty microwave filter according to the solutions, including debugging and design.

[0031] Step 7.5: When the similarity condition is not met, the same method as in Step 6.1 is used to extract the scattering parameters and component process dimensions of the faulty microwave filter from the microwave filter knowledge graph; the scattering parameters and component process dimensions of the faulty microwave filter are combined into a set of input parameters for the diagnosis and repair inversion model.

[0032] Step 7.6: Using the same method as steps 6.2 and 6.3, obtain the inversion result of the faulty microwave filter. The inversion result includes the coupling matrix, self-coupling coefficient, mutual coupling coefficient, and component adjustment of the faulty filter. Store the inversion knowledge triplet of the faulty microwave filter in the Neo4j graph database to obtain the microwave filter knowledge graph after updating the inversion knowledge triplet.

[0033] Step 7.7: In the microwave filter knowledge graph after updating the inversion knowledge triplet, query the coupling matrix and ideal coupling matrix of the faulty filter, calculate the difference matrix between the coupling matrix and the ideal coupling matrix, use the position of the non-zero diagonal element in the difference matrix to represent the position of the faulty resonant unit of the faulty microwave filter, diagnose the faulty resonant unit of the faulty microwave filter, and debug and modify the design scheme for the faulty resonant unit.

[0034] Step 7.8: The commissioning personnel and designers look up the component adjustment amounts of the faulty microwave filter in the updated microwave filter knowledge graph. The commissioning personnel adjust the microwave filter according to the component adjustment amounts, and the designers modify the component dimensions in the design scheme of the faulty microwave filter according to the component adjustment amounts.

[0035] Compared with the prior art, the present invention has the following advantages:

[0036] First, this invention constructs a microwave filter knowledge graph ontology, overcoming the shortcomings of existing technologies in extracting unstructured text data. This is because the unstructured text data across the entire microwave filter cycle often lacks connections, making effective extraction of relationships impossible and resulting in disconnected data. This invention enables the effective acquisition of valuable information from unstructured data, achieving the association and visualization of microwave filter data across its entire lifecycle, and improving data utilization.

[0037] Secondly, this invention integrates diagnostic and repair inversion models into a microwave filter knowledge graph. The knowledge graph, constructed based on full-cycle data, enables diagnostic and repair inversion, using the inversion results to guide debugging and design, overcoming the shortcomings of the microwave filter industry in terms of data resource waste. This allows the invention to fully utilize full-cycle microwave filter data, shortening the response cycle for problem feedback from manufacturing to design stages and improving the efficiency of design and manufacturing iterations. Attached Figure Description

[0038] Figure 1 This is a flowchart of the present invention;

[0039] Figure 2 This is a schematic diagram of the microwave filter knowledge graph ontology of the present invention;

[0040] Figure 3 This is a schematic diagram of the microwave filter knowledge graph fusion diagnosis and repair inversion model of the present invention. Detailed Implementation

[0041] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0042] Reference Figure 1 The implementation steps of the embodiments of the present invention will be described in further detail.

[0043] Step 1: Collect full-cycle unstructured text data of the microwave filter.

[0044] In this embodiment of the invention, a text dataset generated by a full-cycle microwave filter is used as the source of unstructured data. This full-cycle text dataset contains 1,000 unstructured text data entries.

[0045] Step 2: Construct an ontology of microwave filter knowledge graph based on full-cycle data.

[0046] In this embodiment of the invention, 1,000 text data entries contained in the full-cycle text dataset of the microwave filter in step 1 are used as the knowledge source for the microwave filter knowledge graph.

[0047] Since the full lifecycle data of microwave filters involves both the design and manufacturing stages, it encompasses professional knowledge in the field of microwave filters. The full lifecycle data of microwave filters mainly includes filter design documents, manufacturing process records, and diagnostic and maintenance records. This embodiment of the invention constructs an ontology of microwave filter knowledge graphs based on expert knowledge.

[0048] The ontology construction in this embodiment of the invention defines entities and relationships in the microwave filter knowledge graph, providing a standard for subsequent entity extraction and knowledge triple construction.

[0049] Reference Figure 2 In this embodiment of the invention, the entities of the microwave filter knowledge graph are divided into 28 entity types: microwave filter name, unit name, component name, ideal in-band VSWR, ideal center frequency, ideal bandwidth, ideal insertion loss, ideal return loss, ideal out-of-band suppression, ideal coupling matrix, material name, design structure, metal shell size, component size, designer, production date, fault description, fault cause, solution, metal shell process size, component process size, scattering parameters, center frequency, bandwidth, insertion loss, return loss, out-of-band suppression, and operating conditions.

[0050] In this embodiment of the invention, the relationships in the microwave filter knowledge graph are divided into eight types: design indicators, design information, manufacturing information, manufacturing performance, fault information, components, causes, and solutions. Using 28 entity types and 8 relationship types, 28 "head entity-relationship-tail entity" knowledge triples are constructed as shown in Table 1, where each row in Table 1 represents a knowledge triple.

[0051] Table 1. List of 28 types of knowledge triples

[0052]

[0053] Step 3: Generate training and test sets.

[0054] Step 3.1: Irrelevant data removal and long text segmentation are performed sequentially on the samples in the full-cycle unstructured text data of the microwave filter.

[0055] The text data in the embodiments of the present invention is processed as follows:

[0056] The text data is split into lines according to each period, and spaces and symbols other than commas and periods are removed to obtain the processed text data.

[0057] For example, after processing the text data "The test temperature of this microfilter is 0 to 60 degrees Celsius. At this temperature, the bandwidth of the filter is 100MHz." as described above, we get the two lines of text "The test temperature of this microwave filter is 0 to 60 degrees Celsius." and "At this test temperature, the bandwidth of the filter is 100Hz."

[0058] Step 3.2: Randomly select 200 text data from the processed text data to form a sample set, randomly select 140 text data from the sample set to form a training set, and the remaining 60 text data to form a test set.

[0059] Step 3.3: Use BMEO sequence labeling to label the samples in the preprocessed unstructured text data with their corresponding entity types. The first character of each entity is labeled as "B-entity type"; the middle character is labeled as "M-entity type"; the last character is labeled as "E-entity type"; and irrelevant characters are labeled as "O".

[0060] For example, the text data "The production date of the sixth-order filter is October" can be annotated using the BEMO sequence labeling method. The sixth is labeled as "B-microwave filter name", the order, filter, and wave are labeled as "M-microwave filter name", the device is labeled as "E-microwave filter name", the ten is labeled as "B-production date", the month is labeled as "E-production date", and the of, production, production, day, and period are labeled as "O".

[0061] Step 4: Obtain a qualified entity extraction model based on a bidirectional long short-term memory network.

[0062] Step 4.1: Input the text data from the training set into the entity extraction model based on a bidirectional long short-term memory network, and output the predicted labels. Use stochastic gradient descent to iteratively update the parameters in the entity extraction model until the loss function converges, thus obtaining the trained entity extraction model.

[0063] The loss function is as follows:

[0064]

[0065] Where MSE is the loss value between the labels predicted by the entity extraction model and the labels labeled in the training set, n is the number of labels in the training set, and y i The label for the i-th entity in the training set. Let be the label of the i-th predicted label.

[0066] The entity extraction model based on the bidirectional long short-term memory network algorithm is a model constructed using existing technology. For its structure and the setting of initial parameters, please refer to the description by Nie Tongpan et al. in their paper "Knowledge Graph Construction Technology and Application for Fault Diagnosis of Aircraft Power Systems" (Acta Aeronautica Sinica, 2021, 42).

[0067] Step 4.2: Input the test set into the trained entity extraction model, output the entity type label predicted for each text data in the test set, calculate the accuracy of the entity type labels of all text data output by the entity extraction model, and obtain a qualified entity extraction model when the extraction accuracy of the extraction model reaches 80%.

[0068] The accuracy formula is as follows:

[0069]

[0070] Where: P is the accuracy of the entity extraction model in extracting labels, T is the total number of entity extraction labels that match the labeled labels, and F is the total number of entity extraction labels that do not match the labeled labels.

[0071] Step 5: Construct knowledge triples for the microwave filter knowledge graph.

[0072] All text data is input into a qualified entity extraction model, which outputs the entities in the text data and their corresponding entity types. Based on the entity type, the entities in the text data are associated with the "relationships" in the knowledge triples corresponding to that entity type in Table 1, thus constructing knowledge triples for the microwave filter knowledge graph.

[0073] For example, the entities extracted from the text data "The fault of the sixth-order filter is inductor damage." are "sixth-order LC filter" and "inductor". The entity type of "sixth-order LC filter" is "microwave filter name", and the entity type of "inductor damage" is "fault description". The relationship in the triple "microwave filter name-fault information-fault description" in the ontology is "fault information". Using "fault information" to associate entities, the triple "sixth-order filter-fault information-inductor damage" is constructed.

[0074] Step 6: Construct a knowledge graph of microwave filters.

[0075] All knowledge triples are stored in the Neo4j graph database. Using Neo4j graph database generation commands, each knowledge triple's entity is treated as a corresponding node, and the relation is used as an edge connecting the head and tail entity nodes, thus constructing a knowledge graph for the microwave filter and enabling the association of data across the entire microwave filter lifecycle.

[0076] Neo4j is a high-performance, open-source, non-relational graph database developed by Neo4j.

[0077] Step 7: Microwave filter knowledge graph fusion diagnosis and repair inversion model.

[0078] Step 7.1, please refer to Figure 3 Extract all entities of type scattering parameter and component process size from the microwave filter knowledge graph; combine a set of associated scattering parameters and component process sizes into a set of input parameters for a diagnostic and repair inversion model.

[0079] Step 7.2: Select a set of previously unselected diagnostic and repair inversion input parameters. Remove the phase loading from the scattering parameters in the input parameters to obtain the true scattering parameters. Convert the scattering parameters S into the admittance parameters Y according to the following formula:

[0080]

[0081]

[0082]

[0083]

[0084] Among them, Y 11 ,Y 21 ,Y 12 ,Y 22 The admittance parameters are divided into positive input admittance, positive transfer admittance, negative transfer admittance, and negative input admittance, S. 11 ,S 21 ,S 12 ,S 22 These are respectively represented as the forward reflection coefficient, forward transmission coefficient, reverse transmission coefficient, and reverse reflection coefficient in the scattering parameters.

[0085] According to the following formula, the admittance parameter is decomposed into fractions to obtain the poles and residues of the admittance parameter. The poles and residues of the admittance parameter can be solved using a vector fitting algorithm.

[0086]

[0087] Where Y is the admittance parameter, K ∞ λ is a factor related to the finite number of transmission zeros. k The poles of the admittance parameter are s = jω, where ω is the frequency and r is the frequency. 11k r 12k r 21k r 22k These are the residues of the corresponding admittance parameters, and N represents the order of the microwave filter.

[0088] The poles and residues of the scattering parameters are used to obtain the coupling matrix M of the microwave filter using the following formula, based on the coupling matrix synthesis theory.

[0089] M SL =K ∞

[0090] M kk =-λ k

[0091]

[0092]

[0093] Among them, M SL M represents the mutual coupling coefficient between the source and the load. Sk M represents the mutual coupling coefficient between the source and the k-th resonant cavity. Lk M represents the mutual coupling coefficient between the load and the k-th resonant cavity. kk denoted by , where represents the self-coupling coefficient of the k-th resonant cavity, and N represents the order of the microwave filter.

[0094] Step 7.3: Substitute the coupling matrix M and the component process dimensions into the space mapping algorithm, and iterate according to the following formula. When the residual between the coupling matrix and the ideal coupling matrix is ​​less than the threshold, stop the iteration and solve for the component adjustment amount that meets the performance index of the microwave filter.

[0095]

[0096]

[0097]

[0098] ||f (j) ||≤ε

[0099] Among them, B (j) To approximate the Jacobian matrix in the j-th iteration of the space mapping method, B is initially set to the identity matrix, h (j) f is the component adjustment amount in the j-th iteration. (j) Let the residual be the coupling matrix and the ideal coupling matrix at the j-th iteration. Let ε be the component process size at the j-th iteration, and ε be the threshold.

[0100] Step 7.4: The coupling matrix, self-coupling coefficient, mutual coupling coefficient, and component adjustment amount obtained by inversion are correlated with the scattering parameters and component process dimensions in the same input group to construct 5 inversion knowledge triplets, as shown in Table 2.

[0101] Table 2. List of 5 Inversion Knowledge Triplets

[0102]

[0103] Five inversion knowledge triples were imported into the Neo4j graph database for storage, realizing the fusion of microwave filter knowledge graph with microwave filter diagnosis and repair inversion model. Based on the microwave filter commissioning data and design information, effective information was inverted and visualized on the knowledge graph.

[0104] Step 7.5: Determine whether all input parameters for diagnosis and repair inversion have been selected. If yes, proceed to step 8; otherwise, proceed to step 7.2.

[0105] Step 8: Perform diagnosis and repair inversion based on the microwave filter knowledge graph.

[0106] Step 8.1: When a microwave filter malfunctions, the same method as in Steps 3.1 and 5 is used to extract entities from the text data of the malfunctioning microwave filter, obtain the knowledge triplet of the malfunctioning microwave filter, store the knowledge triplet of the malfunctioning microwave filter in the Neo4j graph database, and obtain the microwave filter knowledge graph after updating the knowledge triplet of the malfunctioning microwave filter.

[0107] Step 8.2: Using a similarity algorithm, match the microwave filter name and scattering parameters that are similar to those of the faulty microwave filter in the updated microwave filter knowledge graph.

[0108] Step 8.3: Determine whether the similarity between the microwave filter name in the knowledge graph and the microwave filter name of the faulty microwave filter, and the similarity between the scattering parameters in the microwave filter knowledge graph and the scattering parameters of the faulty microwave filter, satisfy the similarity conditions. If yes, proceed to step 8.4; otherwise, proceed to step 8.5. The similarity algorithm refers to any one of the following algorithms: cosine similarity algorithm and longest common subsequence algorithm.

[0109] The similarity criteria are that the similarity between the microwave filter name in the microwave filter knowledge graph and the microwave filter name of the faulty microwave filter reaches 100%, and the similarity between the scattering parameters in the microwave filter knowledge graph and the scattering parameters of the faulty microwave filter reaches 90%.

[0110] Step 8.4: Use Cypher to query the knowledge graph of the microwave filter and find the solutions corresponding to the microwave filter name and scattering parameters; repair the faulty microwave filter according to the solutions.

[0111] Step 8.5: When the similarity condition is not met, the same method as in step 7.1 is used to extract the scattering parameters and component process dimensions of the faulty microwave filter from the microwave filter knowledge graph, and the scattering parameters and component process dimensions of the faulty microwave filter are combined into a set of input parameters for the diagnosis and repair inversion model.

[0112] Step 8.6: Using the same method as in Step 7, obtain the inversion result of the faulty microwave filter. The inversion result includes the coupling matrix, self-coupling coefficient, mutual coupling coefficient, component size adjustment, and inductance value correction of the faulty filter. Store the inversion knowledge triplet of the faulty microwave filter in the Neo4j graph database to obtain the microwave filter knowledge graph after updating the inversion knowledge triplet.

[0113] Step 8.5: Using Cypher, query the coupling matrix and ideal coupling matrix of the faulty filter in the microwave filter knowledge graph, calculate the difference matrix between the coupling matrix and the ideal coupling matrix. The position of the non-zero diagonal element in the difference matrix represents the position of the faulty resonant unit of the faulty microwave filter. Diagnose the faulty resonant unit of the faulty microwave filter, and debug and modify the design scheme for the faulty resonant unit.

[0114] Step 8.6: Using the Cypher language, query the component adjustment values ​​of the faulty microwave filter in the microwave filter knowledge graph. The debugging personnel adjust the microwave filter according to these values, and the designer modifies the component design dimensions in the faulty microwave filter design scheme based on these adjustments. Through debugging and design modification, the faulty microwave filter is repaired.

Claims

1. A knowledge graph based microwave filter diagnosis and repair inversion method, characterized in that, An ontology of microwave filter knowledge graph is constructed, and a diagnostic and repair inversion model is fused using the microwave filter knowledge graph. Diagnostic and repair inversion is performed based on the microwave filter knowledge graph. The steps of this inversion method include the following: Step 1: Collect unstructured text data generated throughout the entire cycle of the microwave filter; Step 2, construct the ontology of the microwave filter knowledge graph: Based on the unstructured text data generated throughout the entire lifecycle of the microwave filter, which includes microwave filter design documents, manufacturing process records, and diagnostic and maintenance records, 28 entity types and 8 relation types are defined respectively. The 28 entity types and 8 relation types are then constructed into 28 "head entity-relationship-tail entity" knowledge triples. All entity types, relation types, and knowledge triples are combined to form the ontology of the microwave filter knowledge graph. Step 3: Generate training and test sets: Step 3.1: Preprocess the unstructured text data; Step 3.2: Randomly select 25% of the text data from the unstructured text data and label the entity type of each selected text data. Step 3.3: Generate training and testing sets from the labeled text data in an 8:2 ratio; Step 4: Obtain a qualified entity extraction model based on a bidirectional long short-term memory network. Step 4.1: Input the training set into the entity extraction model based on the bidirectional long short-term memory network, and use the stochastic gradient descent method to iteratively update the parameters in the entity extraction model until the loss function converges, thus obtaining the trained entity extraction model. Step 4.2: Input the test set into the trained entity extraction model, output the entity type label predicted for each text data in the test set, calculate the accuracy of the entity type labels of all text data output by the entity extraction model, and obtain a qualified entity extraction model when the extraction accuracy of the extraction model reaches 80%. Step 5, construct a microwave filter knowledge graph: Step 5.1: Input all text data in the unstructured text data into a qualified entity extraction model, and output the entity of each text data and its corresponding entity type. Step 5.2: Associate the entities in each text data according to their entity types and the knowledge triples in the ontology to obtain the knowledge triples of the microwave filter knowledge graph for that text data. Step 5.3: Store all knowledge triples in the Neo4j graph database. Using the Neo4j graph database generation instructions, treat each knowledge triple entity as a corresponding node and the relation as an edge connecting the head entity node and the tail entity node to obtain the microwave filter knowledge graph. Step 6: Utilize the microwave filter knowledge graph to fuse the diagnostic and repair inversion model: Step 6.1: Extract all entities with entity types of scattering parameters and component process dimensions from the microwave filter knowledge graph; combine a set of associated scattering parameters and component process dimensions into a set of input parameters for the diagnostic and repair inversion model; Step 6.2: Select a set of previously unselected diagnostic and repair inversion input parameters; convert the scattering parameters in the selected diagnostic and repair inversion input parameters into admittance parameters, substitute them into the vector fitting algorithm, and obtain the coupling matrix including the coupling coefficient and mutual coupling coefficient according to the coupling matrix synthesis theory. Substituting the component process dimensions and the obtained coupling matrix from the input parameters into the space mapping algorithm, the component adjustment amount that meets the performance index of the microwave filter is obtained; Step 6.3: Correlate the inverted coupling matrix, self-coupling coefficient, mutual coupling coefficient, and component adjustment amount with the scattering parameters and component process dimensions in the same set of input parameters; Five inversion knowledge triples are obtained: "scattering parameters-inversion-coupling matrix", "coupling matrix-including self-coupling coefficient", "coupling matrix-including mutual coupling coefficient", "coupling matrix-inversion-component adjustment amount", and "component process dimensions-inversion-component adjustment amount". These five inversion knowledge triples are stored in the Neo4j graph database to obtain the microwave filter knowledge graph after updating the inversion knowledge triples. Step 6.4: Determine whether all input parameters for diagnosis and repair inversion have been selected. If yes, proceed to step 7; otherwise, proceed to step 6.

2. Step 7: Perform diagnosis and repair inversion based on the microwave filter knowledge graph: Step 7.1: When a microwave filter fails, the same method as in Steps 3.1, 5.1 and 5.2 is used to extract entities from the text data of the failed microwave filter, obtain the knowledge triplet of the failed microwave filter, store the knowledge triplet of the failed microwave filter in the Neo4j graph database, and obtain the microwave filter knowledge graph after updating the knowledge triplet of the failed microwave filter. Step 7.2: Using a similarity algorithm, match the microwave filter name and scattering parameters that are similar to those of the faulty microwave filter in the updated microwave filter knowledge graph. Step 7.3: Determine whether the similarity between the microwave filter name in the knowledge graph and the microwave filter name of the faulty microwave filter, and the similarity between the scattering parameters in the microwave filter knowledge graph and the scattering parameters of the faulty microwave filter, meet the similarity conditions. If so, proceed to step 7.

4. Otherwise, proceed to step 7.5; Step 7.4: The debugging and design personnel search for solutions corresponding to the microwave filter name and scattering parameters in the microwave filter knowledge graph; and repair the faulty microwave filter according to the solutions, including debugging and design. Step 7.5: When the similarity condition is not met, the same method as in Step 6.1 is used to extract the scattering parameters and component process dimensions of the faulty microwave filter from the microwave filter knowledge graph; the scattering parameters and component process dimensions of the faulty microwave filter are combined into a set of input parameters for the diagnosis and repair inversion model. Step 7.6: Using the same method as steps 6.2 and 6.3, obtain the inversion result of the faulty microwave filter. The inversion result includes the coupling matrix, self-coupling coefficient, mutual coupling coefficient, and component adjustment amount of the faulty filter. The inversion knowledge triplet of the faulty microwave filter is stored in the Neo4j graph database to obtain the microwave filter knowledge graph after updating the inversion knowledge triplet. Step 7.7: In the microwave filter knowledge graph after updating the inversion knowledge triplet, query the coupling matrix and ideal coupling matrix of the faulty filter, calculate the difference matrix between the coupling matrix and the ideal coupling matrix, use the position of the non-zero diagonal element in the difference matrix to represent the position of the faulty resonant unit of the faulty microwave filter, diagnose the faulty resonant unit of the faulty microwave filter, and debug and modify the design scheme for the faulty resonant unit. Step 7.8: The commissioning personnel and designers look up the component adjustment amounts of the faulty microwave filter in the updated microwave filter knowledge graph. The commissioning personnel adjust the microwave filter according to the component adjustment amounts, and the designers modify the component dimensions in the design scheme of the faulty microwave filter according to the component adjustment amounts.

2. The knowledge graph based microwave filter diagnosis and cure inversion method of claim 1, wherein, The step 2, which involves constructing 28 "head entity-relationship-tail entity" knowledge triplets for the 28 entity types and 8 relation types, refers to the following: "Microwave filter name," "Unit name," "Component name," "Ideal in-band VSWR," "Ideal center frequency," "Ideal bandwidth," "Ideal insertion loss," "Ideal return loss," "Ideal out-of-band rejection," "Ideal coupling matrix," "Material name," "Design structure," "Metal casing dimensions," "Component dimensions," "Designer," "Production date," "Fault description," "Fault cause," "Solution," "Metal casing process dimensions," "Component process dimensions," and so on. The system incorporates 28 entity types ("scattering parameters", "center frequency", "bandwidth", "insertion loss", "return loss", "out-of-band rejection", and "operating conditions") and 8 relationship types ("design specifications", "design information", "manufacturing information", "manufacturing performance", "fault information", "component", "cause", and "solution") into a structure called "microwave filter name-component-unit name", "unit name-component-element name", "microwave filter name-design specifications-ideal center frequency", "microwave filter name-design specifications-ideal insertion loss", "microwave filter name-design specifications-ideal return loss", and "microwave filter name-design specifications-ideal return loss". Design Specifications - Ideal Out-of-Band Suppression, Microwave Filter Name - Design Specifications - Ideal Coupling Matrix, Microwave Filter Name - Design Information - Material Name, Microwave Filter Name - Design Information - Design Structure, Microwave Filter Name - Design Information - Metal Shell Dimensions, Microwave Filter Name - Design Information - Component Dimensions, Microwave Filter Name - Design Information - Designer, Microwave Filter Name - Fault Information - Fault Description, Fault Description - Cause - Fault Cause, Fault Description - Manufacturing Information - Scattering Parameters, Fault Description - Solution, Microwave Filter Name - Manufacturing Information - Production Date The document contains 28 types of "head entity-relationship-tail entity" knowledge triples, including "microwave filter name-manufacturing information-scattering parameters", "microwave filter name-manufacturing information-operating conditions", "microwave filter name-manufacturing information-metal shell process dimensions", "microwave filter name-manufacturing information-component process dimensions", "scattering parameters-manufacturing information-operating conditions", "scattering parameters-manufacturing performance-center frequency", "scattering parameters-manufacturing performance-bandwidth", "scattering parameters-manufacturing performance-insertion loss", "scattering parameters-manufacturing performance-return loss", "scattering parameters-manufacturing performance-out-of-band suppression", and "component name-manufacturing information-component process dimensions".

3. The knowledge graph based microwave filter diagnosis and cure inversion method of claim 1, wherein, The preprocessing mentioned in step 3.1 refers to sequentially removing irrelevant data and cutting long texts from the samples in the unstructured text data of the entire microwave filter cycle.

4. The knowledge graph based microwave filter diagnosis and cure inversion method of claim 1, wherein, The entity type labeling of each selected text data mentioned in step 3.2 refers to using the BMEO sequence labeling method to label the samples in the preprocessed unstructured text data with their corresponding entity types.

5. The knowledge graph based microwave filter diagnosis and cure inversion method of claim 1, wherein, The loss function described in step 4.1 is as follows: Wherein, MSE is the loss value between the label predicted by the entity extraction model and the label annotated in the training set, n is the number of training set labels, y i is the label annotated in the training set, is the i-th predicted label.

6. The knowledge graph based microwave filter diagnosis and cure inversion method of claim 1, wherein, The accuracy rate mentioned in step 4.2 is obtained by the following formula: Where P is the accuracy of the entity extraction model's label extraction, T is the number of labels output by the entity extraction model that match the labels labeled in the test set, and F is the number of labels output by the entity extraction model that do not match the labels labeled in the test set.

7. The knowledge graph based microwave filter diagnosis and cure inversion method of claim 1, wherein, The similarity conditions mentioned in step 7.3 are: the similarity between the microwave filter name in the microwave filter knowledge graph and the microwave filter name of the faulty microwave filter reaches 100%, and the similarity between the scattering parameters in the microwave filter knowledge graph and the scattering parameters of the faulty microwave filter reaches 90%.

Images