An electromagnetic compatibility fault prediction method based on a knowledge graph

Abstract

The application discloses an electromagnetic compatibility fault prediction method based on a knowledge graph, relates to the technical fields of electromagnetic compatibility fault diagnosis and artificial intelligence, and constructs an electromagnetic compatibility fault knowledge graph through multi-source heterogeneous data, realizes the structured fusion of data, provides nodes and relations with physical meanings for a Bayesian network, makes the target Bayesian network have interpretability and be capable of tracing fault sources, adopts ALS-PSO to optimize the target Bayesian network, can effectively improve the training precision of the target Bayesian network, and thus can make the target Bayesian network after training accurately realize the prediction of electromagnetic compatibility faults.

Description

A knowledge graph-based method for electromagnetic compatibility fault prediction Technical Field

[0001] This application relates to the fields of electromagnetic compatibility fault diagnosis and artificial intelligence technology, and more specifically, to an electromagnetic compatibility fault prediction method based on knowledge graphs. Background Technology

[0002] With the increasing complexity and high integration of electronic devices, electromagnetic compatibility (EMC) issues have become a key factor affecting equipment reliability. Traditional EMC fault prediction mainly relies on expert experience or a single physical model. However, EMC faults are often characterized by nonlinearity, strong coupling, and randomness, making them difficult to accurately describe using simple mathematical models.

[0003] In recent years, data-driven fault prediction methods have gradually emerged. Bayesian networks, as a probabilistic graphical model for handling uncertainties, have been widely used in the field of fault diagnosis. However, existing Bayesian network-based fault prediction methods still suffer from the following technical bottlenecks:

[0004] Network structure construction is difficult and inaccurate: Traditional Bayesian network structure learning is often based on pure statistical methods (such as K2 algorithm and BIC score), ignoring domain knowledge. This results in the constructed network structure not conforming to the physical mechanism of electromagnetic compatibility fault propagation, and it is difficult to automatically extract effective nodes and edges when faced with multi-source heterogeneous data.

[0005] Parameter learning is prone to getting trapped in local optima: Training the conditional probability table (CPT) parameters of a Bayesian network is essentially a high-dimensional nonlinear optimization problem. Traditional gradient descent or standard particle swarm optimization (PSO) methods are prone to converging to local optima when dealing with complex objective functions, and suffer from slow convergence speed and low optimization accuracy.

[0006] Low utilization of multi-source data: EMC-related data typically includes test data, operation logs, and maintenance records, which are often semi-structured or unstructured. Existing technologies struggle to effectively integrate this heterogeneous multi-source information, resulting in incomplete input information for prediction models and impacting prediction accuracy. Summary of the Invention

[0007] This application aims to provide an electromagnetic compatibility fault prediction method based on knowledge graphs, which seeks to solve the problems of unreasonable network structure construction, low parameter optimization accuracy, and insufficient utilization of multi-source data in the prior art.

[0008] This application provides a knowledge graph-based electromagnetic compatibility fault prediction method, including:

[0009] Collect multi-source heterogeneous data related to electromagnetic compatibility; the multi-source heterogeneous data includes electromagnetic compatibility test data, equipment operation logs, and fault repair records.

[0010] Natural language processing algorithms are used to extract entities and relationships from the multi-source heterogeneous data and construct an electromagnetic compatibility fault knowledge graph.

[0011] Based on the entity and relation mapping in the electromagnetic compatibility fault knowledge graph, the basic nodes and initial network structure of the Bayesian network are determined, and the target Bayesian network is obtained.

[0012] The conditional probability parameters of the target Bayesian network are trained using ALS-PSO to determine the trained target Bayesian network.

[0013] The real-time electromagnetic compatibility (EMC) status characteristics of the device under test are obtained, and the real-time EMC status characteristics are input into the trained target Bayesian network for probabilistic inference, and the EMC fault prediction results are output.

[0014] In one possible design approach, based on the entity-relation mapping in the knowledge graph, the basic nodes and initial network structure of the Bayesian network are determined to obtain the target Bayesian network, including:

[0015] The entities in the knowledge graph are mapped to nodes of a Bayesian network;

[0016] Map the relationships between entities in a knowledge graph to directed edges in a Bayesian network;

[0017] For nodes in the knowledge graph that are not explicitly connected but co-occur in physical space, the connection relationships supplemented by human-computer interaction are obtained and mapped as directed edges to form a complete Bayesian network topology, thus obtaining the target Bayesian network.

[0018] In one possible design approach, ALS-PSO is used to train the conditional probability parameters of the target Bayesian network to determine the trained target Bayesian network, including:

[0019] The conditional probability parameters of the target Bayesian network are encoded to initialize the particle swarm.

[0020] For any particle in the particle swarm, the fitness of the particle is obtained according to the knowledge graph, and the particle with the highest fitness is determined as the optimal particle.

[0021] Based on the optimal particle, the particle is adaptively updated using a cross-strategy of historical gene mutation and social information learning based on least squares, resulting in an adaptively updated particle.

[0022] The known optimal solution space is developed by using the polar coordinate search strategy of Levy flight control to obtain the particle after the known optimal solution space is developed.

[0023] A time-response-based boundary search strategy is used to perform a coordinated search on the particles after the known optimal solution space has been developed, to obtain the particles after the coordinated search.

[0024] Determine the total number of training iterations, and determine the training phase based on the total number of training iterations; wherein the training phase includes a training completion phase or a training incomplete phase;

[0025] If the training phase is incomplete, then return to the step of obtaining the optimal particle;

[0026] If the training phase is the training completion phase, then the optimal particle is re-determined based on the particles after the coordinated search, and the parameters in the re-determined optimal particle are used as the final conditional probability parameters of the target Bayesian network to obtain the trained target Bayesian network.

[0027] In one possible design approach, the conditional probability parameters of the target Bayesian network are encoded, and the particle swarm is initialized, including:

[0028] The conditional probability parameters of the target Bayesian network are randomly initialized between the upper and lower bounds and encoded into a vector to obtain particles;

[0029] By repeatedly generating multiple different particles, a particle swarm is obtained.

[0030] In one possible design approach, based on the optimal particle, the particle is adaptively updated using a strategy combining historical gene mutation and social information learning based on least squares, resulting in an adaptively updated particle, including:

[0031] When the total number of training iterations is less than the threshold, the optimal particle is adaptively updated using a social information learning strategy, resulting in the following adaptively updated particle:

[0032] ;

[0033] ;

[0034] Wherein, the threshold number of training iterations is less than the preset maximum number of training iterations. Let m represent the m-th particle during the t-th training session, where m = 1, 2, ..., M, and M represents the total number of particles. This represents the particle after the m-th adaptive update. express The speed during the t-th training process express The speed during the (t+1)th training session Indicates the first learning factor. Indicates the second learning factor. Represents the first random number between (0,1). This represents the second random number between (0,1). Represents the optimal particle. express The historical best value;

[0035] When the total number of training iterations exceeds the threshold, three historical values ​​of particles are randomly selected to obtain the first, second, and third historical values ​​of the particles.

[0036] Based on the first historical value, the second historical value, and the third historical value, the first gene variation control factor and the second gene variation control factor are obtained using the least squares method:

[0037] ;

[0038] ;

[0039] ;

[0040] in, This represents the first gene mutation control factor corresponding to the m-th particle. This represents the second gene mutation control factor corresponding to the m-th particle. Let d represent the d-th dimension parameter of the first historical value, where d = 1, 2, ..., D, and D represents the total dimension of the parameters in the particle. The d-th dimension parameter represents the second historical value. The d-th parameter represents the third historical value. This represents the first training interval coefficient. This represents the second training interval coefficient. This represents the third training interval coefficient. This represents the training interval coefficient for the i-th training interval, where i = 1, 2, 3. This represents the average fitness value of a particle during the t-th training iteration. This represents the average fitness value of a particle during training, corresponding to the i-th historical value, and T represents the preset maximum number of training iterations. The term is a positive number and is set to 0.01;

[0041] The first gene mutation control factor and the second gene mutation control factor are obtained according to the least squares method, and the particle is mutated as follows:

[0042] ;

[0043] in, Let d represent the d-th dimension parameter of the particle after the m-th mutation, and e represent the natural constant;

[0044] Employing a social information learning strategy, we obtain particles after social informatics.

[0045] Adaptive selection and crossover are performed between the mutated particles and the social informatics-derived particles to obtain adaptively updated particles.

[0046] In one possible design approach, a polar coordinate search strategy based on Levy's flight control is used to develop the known optimal solution space for the adaptively updated particles, resulting in particles with the known optimal solution space developed, including:

[0047] Obtain Levi's flight factor ;

[0048] To obtain the polar coordinate parameters:

[0049] ;

[0050] ;

[0051] in, Indicates the polar radius. Represents polar coordinate angles. Represents a random number between (0, 1). This represents the total dimension of the parameters in the particle. This represents the angle control parameter, and is set to 0.005.

[0052] Based on the polar coordinate parameters and the Levy flight factor, the known optimal solution space is expanded for the adaptively updated particles, resulting in the following particles after the known optimal solution space has been expanded:

[0053]

[0054] in, This represents the particle after the nth adaptive update during the t-th training process. This represents a particle after the nth known optimal solution space has been expanded. Represents the worst-case particle. This represents the optimal particle.

[0055] In one possible design approach, a time-response-based boundary search strategy is employed to perform a coordinated search on the particles after the known optimal solution space has been expanded, resulting in particles after the coordinated search, including:

[0056] The time response factor is obtained as follows:

[0057]

[0058] in, Indicates the time response factor. Represents a random number that follows a Gaussian distribution. Represents the sine function. Represents pi (π). This indicates the preset maximum number of training iterations. Represents the total number of training iterations, and cos represents the cosine function. This represents the fourth random number between (0,1). Let represent the fifth random number between (0,1), and let log represent the logarithmic function. Indicates the peak response control coefficient;

[0059] Based on the time response factor and combined with boundary information, a coordinated search is performed on the particles after the known optimal solution space has been expanded, resulting in the following particles after the coordinated search:

[0060]

[0061]

[0062] in, Let d represent the d-th dimension parameter of the particle after the space of the j-th known optimal solution is expanded during the t-th training process. Let d represent the d-th dimension parameter of the particle after the j-th coordination search. Let d represent the boundary information acquisition factor of the d-th dimension, and s represent the order of the coordination search. Indicates the overall coordination search order. Denotes the d-th dimension parameter of the optimal particle. This represents the upper boundary of the d-th dimension parameter.

[0063] In one possible design approach, determining the training phase based on the total number of training iterations includes:

[0064] If the total number of training iterations is greater than the preset maximum number of training iterations, the training phase is determined to be the training completion phase; otherwise, the training phase is determined to be the training incomplete phase.

[0065] In one possible design approach, the real-time electromagnetic compatibility (EMC) status characteristics of the device under test are acquired, and these characteristics are input into a trained target Bayesian network for probabilistic inference. The network then outputs an EMC fault prediction result, including:

[0066] Acquire the real-time electromagnetic compatibility status characteristics of the device under test;

[0067] The real-time electromagnetic compatibility (EMC) state features are input into the trained target Bayesian network, and the joint tree algorithm is used for inference to output the EMC fault prediction results.

[0068] In one possible design approach, the method further includes: generating an electromagnetic compatibility fault detection report based on the electromagnetic compatibility fault prediction results and according to a preset template.

[0069] Beneficial effects:

[0070] This application provides a knowledge graph-based electromagnetic compatibility (EMC) fault prediction method. By constructing an EMC fault knowledge graph through multi-source heterogeneous data, it not only achieves structured data fusion but also provides physically meaningful nodes and relationships for the Bayesian network, making the target Bayesian network interpretable and able to trace the source of the fault. Furthermore, ALS-PSO is used to optimize the target Bayesian network, which can effectively improve the training accuracy of the target Bayesian network, thereby enabling the trained target Bayesian network to accurately predict EMC faults. Attached Figure Description

[0071] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0072] Figure 1 is a flowchart of an electromagnetic compatibility fault prediction method based on knowledge graph proposed in an embodiment of this application;

[0073] Figure 2 is a flowchart of obtaining the trained target Bayesian network according to an embodiment of this application. Detailed Implementation

[0074] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0075] As shown in Figure 1, this application provides an electromagnetic compatibility fault prediction method based on a knowledge graph, including:

[0076] S101. Collect multi-source heterogeneous data related to electromagnetic compatibility; the multi-source heterogeneous data includes electromagnetic compatibility test data, equipment operation logs and fault repair records.

[0077] For example, this can be achieved through test data such as conducted emissions and radiated emissions collected by EMC testing equipment; operational log data such as voltage, current, and temperature uploaded by equipment sensors; and textual description data from historical maintenance records.

[0078] S102. Natural language processing algorithms are used to extract entities and relationships from the multi-source heterogeneous data, and an electromagnetic compatibility fault knowledge graph is constructed.

[0079] Natural language processing algorithms can be used to extract entities and relationships from electromagnetic compatibility (EMC) test data, equipment operation logs, and fault repair records, and to construct an EMC fault knowledge graph. For example, entities from multi-source heterogeneous data can include equipment entities, interference source entities, sensitive equipment entities, fault phenomenon entities, fault cause entities, and / or environmental parameter entities; relationships can include those such as causing, coupled to, belonging to, occurring at, and / or related to. However, it is worth noting that the above entities and relationships are merely illustrative examples in this application's embodiments; in actual applications, the actual data shall prevail.

[0080] S103. Based on the entity and relation mapping in the electromagnetic compatibility fault knowledge graph, determine the basic nodes and initial network structure of the Bayesian network to obtain the target Bayesian network.

[0081] Entities in a knowledge graph (such as interference source entities and sensitive device entities) can be mapped to nodes in a Bayesian network; relationships between entities (such as causes) can be mapped to directed edges.

[0082] Considering the possibility of missing data, for nodes in the knowledge graph that are not explicitly connected but co-occur in physical space (such as two chips on the same circuit board), experts are consulted via a human-computer interaction interface to inquire whether a connection should be established. If the expert confirms, the connection is supplemented by a directed edge, thus forming a complete Bayesian network topology that conforms to physical logic, which is the prototype of the target Bayesian network.

[0083] S104. The conditional probability parameters of the target Bayesian network are trained using ALS-PSO to determine the trained target Bayesian network.

[0084] After the structure of a Bayesian network is determined, the conditional probability parameters (CPTs) between nodes need to be trained. Existing technologies typically employ particle swarm optimization (PSO) or gradient descent algorithms for training, which often leads to getting trapped in local optima and failing to find the optimal solution in the high-dimensional solution space. Consequently, the trained Target Bayesian network cannot accurately predict electromagnetic compatibility (EMC) faults. Therefore, this application provides an ALS-PSO (Adaptive Least Squares Particle Swarm Optimization) algorithm to train the Target Bayesian network more accurately, enabling it to perform fault reasoning.

[0085] S105. Obtain the real-time electromagnetic compatibility status characteristics of the device under test, and input the real-time electromagnetic compatibility status characteristics into the trained target Bayesian network for probabilistic inference, and output the electromagnetic compatibility fault prediction result.

[0086] For example, the real-time electromagnetic compatibility (EMC) status characteristics of the device under test (DUT) can be obtained (such as real-time spectrum data, operating voltage, etc.). These characteristics are then input into a trained target Bayesian network. A joint tree algorithm is used for efficient probabilistic inference to calculate the posterior probability of each fault node, and the fault type with the highest probability is output as the prediction result.

[0087] This application's embodiments collect electromagnetic compatibility test data, equipment operation logs, and fault repair records, and construct a knowledge graph using natural language processing algorithms. Compared to methods relying solely on single test data, this invention fully utilizes the implicit fault mechanisms and expert experience contained in unstructured text (such as logs and repair records). The construction of the knowledge graph not only achieves structured data fusion but also provides Bayesian networks with physically meaningful nodes and relationships, making the prediction model interpretable and enabling the tracing of fault sources.

[0088] In one possible design approach, based on the entity-relation mapping in the knowledge graph, the basic nodes and initial network structure of the Bayesian network are determined to obtain the target Bayesian network, including:

[0089] The entities in the knowledge graph are mapped to nodes of a Bayesian network;

[0090] Map the relationships between entities in a knowledge graph to directed edges in a Bayesian network;

[0091] For nodes in the knowledge graph that are not explicitly connected but co-occur in physical space, the connection relationships supplemented by human-computer interaction are obtained and mapped as directed edges to form a complete Bayesian network topology, thus obtaining the target Bayesian network.

[0092] This application's embodiments construct the initial structure of the Bayesian network based on the entity-relation mapping in the knowledge graph, avoiding the inefficiency and structural inconsistencies caused by blind searching in traditional pure data-driven methods. In particular, for nodes in the graph that are not explicitly connected but co-occur in physical space, connections are supplemented through human-computer interaction, further improving the network topology. This ensures that the structure of the Bayesian network conforms to the physical laws of electromagnetic compatibility fault propagation, enhancing the fundamental rationality of the model.

[0093] As shown in Figure 2, the conditional probability parameters of the target Bayesian network are trained using ALS-PSO to determine the trained target Bayesian network, including:

[0094] S201. Encode the conditional probability parameters of the target Bayesian network and initialize the particle swarm;

[0095] S202. For any particle in the particle swarm, obtain the fitness of the particle according to the knowledge graph, and determine the particle with the highest fitness as the optimal particle.

[0096] For example, the log-likelihood value of the target Bayesian network can be calculated as the fitness, and the particle with the highest fitness can be determined as the optimal particle. Fitness calculation is a relatively conventional technique for training Bayesian networks, and existing technologies can be used; therefore, it will not be elaborated further in this application.

[0097] S203. Based on the optimal particle, the particle is adaptively updated using a cross-strategy of historical gene mutation and social information learning based on least squares, to obtain the adaptively updated particle.

[0098] S204. Using the polar coordinate search strategy of Levy flight control, the known optimal solution space is developed for the adaptively updated particles to obtain the particles after the known optimal solution space has been developed.

[0099] S205. A time-response-based boundary search strategy is used to perform a coordinated search on the particles after the known optimal solution space has been developed, so as to obtain the particles after the coordinated search.

[0100] S206. Determine the total number of training iterations, and determine the training phase based on the total number of training iterations; wherein the training phase includes a training completion phase or a training incomplete phase;

[0101] S207. If the training phase is incomplete, then return to the step of obtaining the optimal particle.

[0102] S208. If the training phase is the training completion phase, then the optimal particle is re-determined based on the particles after the coordinated search, and the parameters in the re-determined optimal particle are used as the final conditional probability parameters of the target Bayesian network to obtain the trained target Bayesian network.

[0103] In one possible design approach, the conditional probability parameters of the target Bayesian network are encoded, and the particle swarm is initialized, including:

[0104] The conditional probability parameters of the target Bayesian network are randomly initialized between the upper and lower bounds and encoded into a vector to obtain particles;

[0105] By repeatedly generating multiple different particles, a particle swarm is obtained.

[0106] In one possible design approach, based on the optimal particle, the particle is adaptively updated using a strategy combining historical gene mutation and social information learning based on least squares, resulting in an adaptively updated particle, including:

[0107] When the total number of training iterations is less than the threshold (which can be set to T / 4, where T can be a multiple of 100), the optimal particle is adaptively updated using a social information learning strategy, resulting in the following adaptively updated particle:

[0108] ;

[0109] ;

[0110] Wherein, the threshold number of training iterations is less than the preset maximum number of training iterations. Let m represent the m-th particle during the t-th training session, where m = 1, 2, ..., M, and M represents the total number of particles. This represents the particle after the m-th adaptive update. express The speed during the t-th training process express The speed during the (t+1)th training session This represents the first learning factor, which can be set to 0.75; This represents the second learning factor, which can be set to 1.5; Represents the first random number between (0,1). This represents the second random number between (0,1). Represents the optimal particle. express The historical best value;

[0111] When the total number of training iterations exceeds the threshold, three historical values ​​of particles are randomly selected to obtain the first, second, and third historical values ​​of the particles.

[0112] Based on the first historical value, the second historical value, and the third historical value, the first gene variation control factor and the second gene variation control factor are obtained using the least squares method:

[0113] ;

[0114] ;

[0115] ;

[0116] in, This represents the first gene mutation control factor corresponding to the m-th particle. This represents the second gene mutation control factor corresponding to the m-th particle. Let d represent the d-th dimension parameter of the first historical value, where d = 1, 2, ..., D, and D represents the total dimension of the parameters in the particle. The d-th dimension parameter represents the second historical value. The d-th parameter represents the third historical value. This represents the first training interval coefficient. This represents the second training interval coefficient. This represents the third training interval coefficient. This represents the training interval coefficient for the i-th training interval, where i = 1, 2, 3. This represents the average fitness value of a particle during the t-th training iteration. This represents the average fitness value of a particle during training, corresponding to the i-th historical value, and T represents the preset maximum number of training iterations. The term is a positive number and is set to 0.01;

[0117] The first gene mutation control factor and the second gene mutation control factor are obtained according to the least squares method, and the particle is mutated as follows:

[0118] ;

[0119] in, Let d represent the d-th dimension parameter of the particle after the m-th mutation, and e represent the natural constant;

[0120] Employing a social information learning strategy, we obtain particles after social informatics.

[0121] Adaptive selection and crossover are performed between the mutated particles and the social informatics-derived particles to obtain adaptively updated particles.

[0122] Starting with the first-dimensional parameter of the particle, the corresponding parameters from the mutated particle and the social informatics-adapted particle are replaced. The parameter that increases the fitness of the replaced particle is selected as the parameter for the adaptively updated particle. After traversing all dimensions of the particle, the adaptively updated particle is obtained. For example, replacing the first-dimensional parameter of the mutated particle with the first-dimensional parameter yields the first-replaced particle; replacing the first-dimensional parameter of the social informatics-adapted particle with the first-dimensional parameter yields the second-replaced particle. If the fitness of the first-replaced particle is greater than that of the second-replaced particle, then the first-dimensional parameter of the mutated particle is used to replace the first-dimensional parameter of the second-replaced particle. This process is repeated for each dimension of the particle, and after traversing all dimensions, the adaptively updated particle is obtained.

[0123] In the early stages of training, a social information learning strategy is employed to accelerate particle convergence. In the later stages, a least-squares-based historical gene mutation strategy is introduced, utilizing the position information and fitness differences of historical particles to calculate mutation control factors, effectively maintaining population diversity and preventing the algorithm from getting trapped in local optima. The introduction of adaptive selection crossover effectively improves the training speed of the algorithm, balancing training speed and training quality.

[0124] In one possible design approach, a polar coordinate search strategy based on Levy's flight control is used to develop the known optimal solution space for the adaptively updated particles, resulting in particles with the known optimal solution space developed, including:

[0125] Obtain Levi's flight factor ;

[0126] To obtain the polar coordinate parameters:

[0127] ;

[0128] ;

[0129] in, Indicates the polar radius. Represents polar coordinate angles. Represents a random number between (0, 1). This represents the total dimension of the parameters in the particle. This represents the angle control parameter, and is set to 0.005.

[0130] Based on the polar coordinate parameters and the Levy flight factor, the known optimal solution space is expanded for the adaptively updated particles, resulting in the following particles after the known optimal solution space has been expanded:

[0131]

[0132] in, This represents the particle after the nth adaptive update during the t-th training process. This represents a particle after the nth known optimal solution space has been expanded. Represents the worst-case particle. This represents the optimal particle.

[0133] The polar coordinate search strategy of the Levy flight control described above allows the adaptively updated particles to search around the optimal particle. This allows for some positional transformation in the early stages of the algorithm, completely avoiding getting trapped in local optima. While moving towards the optimal particle, it moves away from the worst particle (i.e., the particle with the lowest fitness). Utilizing the long jump characteristic of Levy flight combined with polar coordinate transformation, it performs in-depth exploration within the known optimal solution space. This non-linear search method can escape local limitations and enhance global optimization capabilities. In the later stages of the algorithm, it gradually refines the local search around the optimal particle, ensuring training accuracy.

[0134] In one possible design approach, a time-response-based boundary search strategy is employed to perform a coordinated search on the particles after the known optimal solution space has been expanded, resulting in particles after the coordinated search, including:

[0135] The time response factor is obtained as follows:

[0136]

[0137] in, Indicates the time response factor. Represents a random number that follows a Gaussian distribution. Represents the sine function. Represents pi (π). This indicates the preset maximum number of training iterations. Represents the total number of training iterations, and cos represents the cosine function. This represents the fourth random number between (0,1). Let represent the fifth random number between (0,1), and let log represent the logarithmic function. This represents the peak response control coefficient, which can be set to a constant between 0.75 and 0.9.

[0138] Based on the time response factor and combined with boundary information, a coordinated search is performed on the particles after the known optimal solution space has been expanded, resulting in the following particles after the coordinated search:

[0139]

[0140]

[0141] in, Let d represent the d-th dimension parameter of the particle after the space of the j-th known optimal solution is expanded during the t-th training process. Let d represent the d-th dimension parameter of the particle after the j-th coordination search. Let d represent the boundary information acquisition factor of the d-th dimension, and s represent the order of the coordination search. This indicates the overall coordination search order, which can be set to 4; Denotes the d-th dimension parameter of the optimal particle. This represents the upper bound of the d-th dimension parameter, i.e., the upper limit.

[0142] The aforementioned time-response-based boundary search strategy introduces a dynamic response factor that changes over time. It adjusts the search step size and direction based on the current training progress and boundary information, enabling the algorithm to have a greater global search capability in the early and mid-stages. This effectively avoids the algorithm getting stuck in local optima. At the same time, the algorithm will gradually explore the neighborhood in the later stages to ensure convergence accuracy.

[0143] Optionally, to accelerate training, simulated annealing can be used to control the time-response-based boundary search strategy.

[0144] In one possible design approach, determining the training phase based on the total number of training iterations includes:

[0145] If the total number of training iterations is greater than the preset maximum number of training iterations, the training phase is determined to be the training completion phase; otherwise, the training phase is determined to be the training incomplete phase.

[0146] In one possible design approach, the real-time electromagnetic compatibility (EMC) status characteristics of the device under test are acquired, and these characteristics are input into a trained target Bayesian network for probabilistic inference. The network then outputs an EMC fault prediction result, including:

[0147] Acquire the real-time electromagnetic compatibility status characteristics of the device under test;

[0148] The real-time electromagnetic compatibility (EMC) state features are input into the trained target Bayesian network, and the joint tree algorithm is used for inference to output the EMC fault prediction results.

[0149] For example, a trained target Bayesian network can be used for real-time prediction.

[0150] Collect real-time electromagnetic compatibility status characteristics of the device under test, such as the current interference voltage value of the power port and the current operating temperature.

[0151] These real-time electromagnetic compatibility status features are used as evidence input into the corresponding nodes of the Bayesian network, and the joint tree algorithm is used for inference to calculate the posterior probability of faulty nodes (such as CPU reset, communication packet loss).

[0152] Output prediction results: For example, the probability of CPU reset is 85% and the probability of communication packet loss is 60%, which can be used to issue early warnings.

[0153] In one possible design approach, the method further includes: generating an electromagnetic compatibility fault detection report based on the electromagnetic compatibility fault prediction results and according to a preset template.

[0154] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0155] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, apparatus, electronic devices, and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal equipment to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal equipment, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0156] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more block diagrams.

[0157] These computer program instructions may also be loaded onto a computer or other programmable data processing terminal equipment to cause a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable terminal equipment, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0158] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0159] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0160] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A knowledge graph-based method for predicting electromagnetic compatibility faults, characterized in that, include: Collect multi-source heterogeneous data related to electromagnetic compatibility; the multi-source heterogeneous data includes electromagnetic compatibility test data, equipment operation logs, and fault repair records. Natural language processing algorithms are used to extract entities and relationships from the multi-source heterogeneous data, and an electromagnetic compatibility (EMC) fault knowledge graph is constructed. Based on the entity and relationship mapping in the EMC fault knowledge graph, the basic nodes and initial network structure of the Bayesian network are determined to obtain the target Bayesian network. The conditional probability parameters of the target Bayesian network are trained using ALS-PSO to determine the trained target Bayesian network. The real-time EMC status features of the device under test are obtained, and the real-time EMC status features are input into the trained target Bayesian network for probabilistic inference to output the EMC fault prediction result. The conditional probability parameters of the target Bayesian network are trained using ALS-PSO to determine the trained target Bayesian network. This process includes: encoding the conditional probability parameters of the target Bayesian network and initializing a particle swarm; for any particle in the swarm, obtaining the fitness of the particle based on the knowledge graph, and determining the particle with the highest fitness as the optimal particle; using the optimal particle as a basis, adaptively updating the particle using a least-squares-based historical gene mutation and social information learning cross-strategy to obtain the adaptively updated particle; using a Levy flight control polar coordinate search strategy to develop the known optimal solution space of the adaptively updated particle to obtain the particle after the known optimal solution space has been developed; and using a time-response-based boundary search strategy to further develop the known optimal solution space. After the optimal solution space is developed, the particles are subjected to a coordinated search to obtain particles after the coordinated search; the total number of training iterations is determined, and the training stage is determined based on the total number of training iterations; wherein, the training stage includes a training completion stage or a training incomplete stage; if the training stage is a training incomplete stage, the step of obtaining the optimal particle is returned; if the training stage is a training completion stage, the optimal particle is re-determined based on the particles after the coordinated search, and the parameters in the re-determined optimal particle are used as the final conditional probability parameters of the target Bayesian network to obtain the trained target Bayesian network; a time-response-based boundary search strategy is used to perform a coordinated search on the particles after the known optimal solution space is developed to obtain particles after the coordinated search, including: obtaining the time response factor as: ;in, Indicates the time response factor. Represents a random number that follows a Gaussian distribution. Represents the sine function. Represents pi (π). This indicates the preset maximum number of training iterations. Represents the total number of training iterations, and cos represents the cosine function. This represents the fourth random number between (0,1). Let represent the fifth random number between (0,1), and let log represent the logarithmic function. This represents the response peak control coefficient; based on the time response factor and combined with boundary information, a coordinated search is performed on the particles after the known optimal solution space has been developed, resulting in the following particles after the coordinated search: ； ;in, Let d represent the d-th dimension parameter of the particle after the space of the j-th known optimal solution is expanded during the t-th training process. Let d represent the d-th dimension parameter of the particle after the j-th coordination search. Let d represent the boundary information acquisition factor of the d-th dimension, and s represent the order of the coordination search. Indicates the overall coordination search order. Denotes the d-th dimension parameter of the optimal particle. This represents the upper boundary of the d-th dimension parameter.

2. The knowledge graph-based electromagnetic compatibility fault prediction method according to claim 1, characterized in that, Based on the entity-relation mapping in the knowledge graph, the basic nodes and initial network structure of the Bayesian network are determined to obtain the target Bayesian network. This includes: mapping entities in the knowledge graph to nodes of the Bayesian network; mapping relations between entities in the knowledge graph to directed edges of the Bayesian network; and for nodes in the knowledge graph that are not explicitly connected but co-occur in physical space, obtaining the connection relations supplemented by human-computer interaction and mapping them to directed edges to form a complete topology of the Bayesian network, thus obtaining the target Bayesian network.

3. The knowledge graph-based electromagnetic compatibility fault prediction method according to claim 1, characterized in that, Encoding the conditional probability parameters of the target Bayesian network and initializing the particle swarm includes: randomly initializing the conditional probability parameters of the target Bayesian network between the upper and lower bounds and encoding them into vectors to obtain particles; repeatedly generating multiple different particles to obtain the particle swarm.

4. The knowledge graph-based electromagnetic compatibility fault prediction method according to claim 1, characterized in that, Based on the optimal particle, an adaptive update is performed on the particle using a cross-strategy of historical gene mutation and social information learning based on least squares, resulting in an adaptively updated particle. This includes: when the total number of training iterations is less than a threshold, the particle is adaptively updated using a social information learning strategy based on the optimal particle, resulting in the following adaptively updated particle: ； Wherein, the threshold number of training iterations is less than the preset maximum number of training iterations. Let m represent the m-th particle during the t-th training session, where m = 1, 2, ..., M, and M represents the total number of particles. This represents the particle after the m-th adaptive update. express The speed during the t-th training process express The speed during the (t+1)th training session Indicates the first learning factor. Indicates the second learning factor. Represents the first random number between (0,1). This represents the second random number between (0,1). Represents the optimal particle. express The historical optimal value; when the total number of training iterations exceeds the threshold, three historical values ​​of particles are randomly selected to obtain the first, second, and third historical values ​​corresponding to the particles; based on the first, second, and third historical values, the first gene mutation control factor and the second gene mutation control factor are obtained using the least squares method: ； ； ;in, This represents the first gene mutation control factor corresponding to the m-th particle. This represents the second gene mutation control factor corresponding to the m-th particle. Let d represent the d-th dimension parameter of the first historical value, where d = 1, 2, ..., D, and D represents the total dimension of the parameters in the particle. This represents the d-th dimension parameter of the second historical value. The d-th parameter represents the third historical value. This represents the first training interval coefficient. This represents the second training interval coefficient. This represents the third training interval coefficient. This represents the training interval coefficient for the i-th training interval, where i = 1, 2, 3. This represents the average fitness value of a particle during the t-th training iteration. This represents the average fitness value of a particle during training, corresponding to the i-th historical value, and T represents the preset maximum number of training iterations. The term is a positive number and is set to 0.01; the first gene mutation control factor and the second gene mutation control factor are obtained according to the least squares method, and the particle is mutated as follows: ;in, Let d represent the d-th dimension parameter of the m-th mutated particle, and e represent the natural constant. A social information learning strategy is adopted to obtain the social information-based particles. Adaptive selection and crossover are performed between the mutated particles and the social information-based particles to obtain the adaptively updated particles.

5. The knowledge graph-based electromagnetic compatibility fault prediction method according to claim 1, characterized in that, The known optimal solution space of the adaptively updated particles is developed using the polar coordinate search strategy of Lévy flight control, resulting in particles with the known optimal solution space developed, including: obtaining the Lévy flight factor as... To obtain the polar coordinate parameters: ； ;in, Indicates the polar radius. Represents the angle in polar coordinates. Represents a random number between (0,1). This represents the total dimension of the parameters in the particle. The angle control parameter is set to 0.

005. Based on the polar coordinate parameters and the Levy flight factor, the known optimal solution space is developed for the adaptively updated particles, resulting in the following particles after the known optimal solution space development: ;in, This represents the particle after the nth adaptive update during the t-th training process. This represents a particle after the nth known optimal solution space has been expanded. Represents the worst-case particle. This represents the optimal particle.

6. The knowledge graph-based electromagnetic compatibility fault prediction method according to claim 1, characterized in that, Determining the training stage based on the total number of training iterations includes: if the total number of training iterations is greater than the preset maximum number of training iterations, then the training stage is determined to be the training completion stage; otherwise, the training stage is determined to be the training incomplete stage.

7. The knowledge graph-based electromagnetic compatibility fault prediction method according to claim 1, characterized in that, The method involves acquiring the real-time electromagnetic compatibility (EMC) status characteristics of the device under test (DUT), inputting these characteristics into a trained target Bayesian network for probabilistic inference, and outputting an EMC fault prediction result. This includes: acquiring the real-time EMC status characteristics of the DUT; inputting these characteristics into a trained target Bayesian network, performing inference using a joint tree algorithm, and outputting an EMC fault prediction result.

8. The knowledge graph-based electromagnetic compatibility fault prediction method according to claim 1, characterized in that, Also includes: Based on the electromagnetic compatibility fault prediction results, an electromagnetic compatibility fault detection report is generated according to a preset template.

Images