Knowledge graph-based device fault diagnosis method and system

Abstract

The application provides a device fault diagnosis method and system based on a knowledge graph. The method comprises: reconstructing a preset initial graph using state adjustment parameters, including correcting fault determination thresholds and adjusting fault propagation links, storing the state adjustment parameters in association with the reconstructed graph, and obtaining a dynamic graph incorporating state adjustment parameters; mapping real-time alarms of a target device to the dynamic graph, retrieving a fault correlation subgraph in the dynamic graph, constructing entity features based on state adjustment parameters, calculating propagation paths using graph attention networks to fuse entity features, and generating an initial candidate fault root cause set; further expanding the root cause set, sorting the root cause set, and determining and outputting the fault cause at the minimum maintenance unit level according to the sorting result. The diagnosis method can determine fault root causes, and small sample faults can also be diagnosed.

Description

Technical Field

[0001] This invention relates to the application of knowledge graphs in equipment fault diagnosis, and more particularly to a method and system for equipment fault diagnosis based on knowledge graphs. Background Technology

[0002] In related technologies, knowledge graphs can currently be used in the field of equipment fault diagnosis, enabling the structured storage and associated retrieval of basic fault knowledge.

[0003] Existing knowledge graph-based equipment fault diagnosis methods often employ a fixed ontology architecture to construct a static fault knowledge graph, combining simple rule matching or basic graph retrieval to complete fault determination. This approach fails to adapt to the impact of sudden load changes and environmental parameter fluctuations in mining equipment, easily leading to misjudgments under normal operating conditions and missed diagnoses of abnormal faults. The diagnostic accuracy continuously decreases with equipment operating time. To address the adaptability issues of static graphs, some existing technologies have begun to introduce dynamic threshold adjustments or perform simple adaptive corrections to graph edge weights. However, these solutions only address the numerical adjustment of a single parameter, failing to consider the spatiotemporal characteristics of fault propagation itself. Furthermore, under scenarios of strong operating condition fluctuations and deep equipment aging, the misjudgment rate of the fault propagation chain remains high.

[0004] Secondly, mining equipment is an operating system with multiple components working together. Existing traditional diagnostic methods cannot determine the root cause of primary and secondary faults, leading to low maintenance efficiency and over-maintenance. Moreover, it is difficult to collect fault samples of mining equipment, and reliable diagnostic results cannot be obtained for small sample faults and zero sample unknown faults. Summary of the Invention

[0005] The purpose of this invention is to provide a knowledge graph-based equipment fault diagnosis method and system, aiming to solve the technical problems of poor adaptability to operating conditions, inaccurate localization of the root causes of coupled faults, and inability to diagnose small sample faults in existing diagnostic methods.

[0006] To achieve the above objectives, the present invention provides the following technical solution.

[0007] In a first aspect, an embodiment of the present invention provides a device fault diagnosis method based on knowledge graphs, comprising:

[0008] The real-time operating conditions of the target equipment are obtained, the steady-state characteristics of the real-time operating conditions are extracted, and the state adjustment parameters of the target equipment are calculated.

[0009] A preset initial map is reconstructed using state adjustment parameters. The reconstruction includes correcting the fault determination threshold and adjusting the fault propagation path. The state adjustment parameters are then associated with the reconstructed map to obtain a dynamic map incorporating the state adjustment parameters. The adjustment of the fault propagation path includes intensity correction and time delay correction.

[0010] The real-time alarms of the target device are mapped to the dynamic graph, and the fault association subgraph is retrieved in the dynamic graph. After the fault association subgraph is topologically optimized, entity features are constructed based on state adjustment parameters. The propagation path is calculated by fusing entity features using a graph attention network to generate an initial candidate fault root cause set.

[0011] A pre-constructed topology transfer model including components and mechanisms is adopted. When there are insufficient target fault samples, the mechanism similarity and topology similarity are corrected by state adjustment parameters. The corrected similarity is used to establish a cross-entity mapping relationship of the graph, and the candidate fault root cause set is expanded through transfer reasoning.

[0012] The candidate fault root cause set is sorted, and the fault cause at the smallest maintenance unit level is determined and output based on the sorting result.

[0013] As a further limitation of this application, the state adjustment parameters include a drift coefficient and a decay coefficient;

[0014] in:

[0015] The drift coefficient is used to characterize the degree to which the current operating condition deviates from the reference operating condition;

[0016] The attenuation coefficient is used to characterize the aging degree of the target equipment.

[0017] As a further limitation of this application, in the reconstruction of the initial map, the step of correcting the fault determination threshold includes:

[0018] Obtain the static fault determination benchmark threshold ;

[0019] Determining the working condition penalty factor based on the drift coefficient , This is the preset drift upper limit cutoff value. This is the drift coefficient;

[0020] The attenuation coefficient is used as the aging correction coefficient. ;

[0021] Calculate the dynamic fault determination threshold Represented as:

[0022]

[0023] In the formula, The drift coefficient, Indicates the baseline threshold Indicates the robustness coefficient under operating conditions. The aging sensitivity coefficient, and satisfying the following conditions. Not lower than the preset physical lower limit threshold .

[0024] As a further limitation of this application, in the reconstruction of the initial spectrum, the intensity correction includes:

[0025] Obtain the original causal association weights between any two related entities in the initial graph. ;

[0026] Using the operating condition disturbance sensitivity coefficient and the aging conduction sensitivity coefficient to weight After correction, adaptive causal association weights are obtained. , represented as: ;in, This represents the sensitivity coefficient to operating disturbances. Indicates the aging conduction sensitivity coefficient; weighting Used to characterize the correlation strength of causal relationships in fault propagation.

[0027] As a further limitation of this application, in the reconstruction of the initial map, the time delay correction includes:

[0028] Obtain the standard propagation delay t0 of the corresponding propagation link in the initial map;

[0029] Based on the drift coefficient and attenuation coefficient, t0 is corrected to obtain the real-time propagation delay t. dyn , represented as:

[0030]

[0031] Where k3 is the operating condition delay coefficient, k4 is the aging delay coefficient, and t dyn Not lower than the preset minimum delay limit t min .

[0032] As a further limitation of this application, the topology optimization of the fault association subgraph includes:

[0033] According to weight With delay Calculate the real-time propagation reachability probability p of the corresponding link. ij , represented as:

[0034]

[0035] in, This is the time delay attenuation coefficient;

[0036] The weakly associated links are temporarily removed from the faulty association subgraph, and these weakly associated links are reached through real-time propagation of the reachability probability p. ij Determined if the value is below a preset reachable threshold;

[0037] If the propagation path is broken due to the removal of weak links, an alternative path is retrieved within a preset number of hops and added to the fault association subgraph as an implicit link.

[0038] As a further limitation of this application, in the calculation of the propagation path, the attention weights between any adjacent entity nodes are determined in the following manner:

[0039] Construct the feature vectors of each entity node ;

[0040] For the directed edge of fault propagation from component entity i to fault phenomenon entity j within the fault association subgraph, calculate the initial attention score. , represented as:

[0041]

[0042] Where W is the shared linear transformation matrix, Let T be the trainable attention parameter vector, and let T denote the transpose of the vector. This represents vector concatenation, with LeakyReLU as the activation function; Let i be the feature vector of entity i. Let j be the feature vector of entity j;

[0043] For all outgoing fault propagation directed edges of the same component entity node i, the attention weights between corresponding adjacent entity nodes are obtained through Softmax normalized exponential mapping. , represented as:

[0044]

[0045] In the formula, Let k be the set of neighboring nodes of node i in the fault association subgraph; k represents the node index. Let represent the attention score between node i and any of its neighboring nodes k.

[0046] As a further limitation of this application, the topology migration model is a two-layer topology migration framework, and the construction method of the two-layer topology migration framework includes:

[0047] Construct a component topology layer, which uses the physical disassembly structure of the equipment as its skeleton, with nodes representing component entities and edges representing the assembly connection relationships and power transmission relationships between components;

[0048] A fault mechanism layer is constructed, which uses the physical mechanism of fault as the skeleton, with nodes representing fault mechanism entities and edges representing the causal evolution relationship between mechanisms;

[0049] A cross-layer mapping edge is established between the component topology layer and the fault mechanism layer, and the cross-layer mapping edge connects the component entity and the fault mechanism entity that the component may fail to.

[0050] The mechanistic similarity and topological similarity are corrected by using state adjustment parameters. Based on the corrected similarity, a cross-entity mapping relationship of the graph is established. Transfer reasoning is performed for small sample faults and zero sample faults respectively to expand the candidate fault root cause set.

[0051] As a further limitation of this application, the sorting of the candidate fault root cause set, and the determination and output of the fault cause at the minimum maintenance unit level based on the sorting result, includes:

[0052] For each candidate root cause entity in the root cause set, obtain the aggregated propagation path score of the entity output by the graph attention network after multiple rounds of message passing;

[0053] For each candidate root cause entity, its potential failure occurrence time is estimated by adding the corresponding real-time propagation delay to obtain the predicted occurrence time of the phenomenon; the predicted occurrence time of the phenomenon is compared with the first alarm timestamp of each fault phenomenon node to calculate the causal direction consistency score of the entity.

[0054] The overall confidence level of the entity is obtained by weighted summation of the aggregated propagation path score, mechanism similarity, topological similarity, and causal direction consistency score.

[0055] Sort the results from highest to lowest overall confidence level, determine the root cause of the target fault based on the sorting results, and output the result.

[0056] Secondly, another embodiment of the present invention provides a knowledge graph-based device fault diagnosis system, comprising the following modules:

[0057] The state perception module is used to acquire the real-time operating conditions of the target device, extract the steady-state characteristics of the real-time operating conditions, and calculate the state adjustment parameters of the target device.

[0058] The graph reconstruction module is used to reconstruct a preset initial graph using state adjustment parameters. The reconstruction includes correcting the fault determination threshold and adjusting the fault propagation link. The state adjustment parameters are associated with the reconstructed graph to obtain a dynamic graph with the state adjustment parameters. The adjustment of the fault propagation link includes intensity correction and time delay correction.

[0059] The path reasoning module is used to map the real-time alarms of the target device to the dynamic graph, retrieve the fault association subgraph in the dynamic graph, perform topology optimization on the fault association subgraph, construct entity features based on state adjustment parameters, use graph attention network to fuse entity features to calculate the propagation path, and generate an initial candidate fault root cause set.

[0060] The migration expansion module is used to adopt a pre-built topology migration model that includes components and mechanisms. When there are insufficient target fault samples, the mechanism similarity and topology similarity are corrected by state adjustment parameters. The corrected similarity is used to establish a cross-entity mapping relationship of the graph, and the candidate fault root cause set is expanded through migration reasoning.

[0061] The sorting output module is used to sort the candidate fault root cause set, determine and output the fault cause at the smallest maintenance unit level based on the sorting result.

[0062] Compared with the prior art, the beneficial effects of the present invention are:

[0063] 1. The initial graph's judgment threshold and fault propagation link attributes are globally reconstructed by adjusting state parameters, and topology optimization is performed on the subgraph during the inference phase. Among them, intensity correction and time delay correction realize the adaptive link attributes, and topology optimization realizes noise filtering and path completion of the inference subgraph, which significantly improves the diagnostic reliability under strong operating condition fluctuations and deep aging scenarios.

[0064] 2. This invention embeds state adjustment parameters into the node features of a graph attention network and combines a dual causal test of alarm timestamp and real-time propagation delay to calculate the propagation path, which can distinguish between primary and secondary faults;

[0065] 3. By constructing a two-layer transfer framework consisting of a component topology layer and a fault mechanism layer, and using state adjustment parameters to drive similarity fusion weights to perform cross-layer mapping transfer reasoning, the reasoning link can be completed based on the dual similarity of mechanism and topology when there are insufficient fault samples, thus realizing the diagnosis of small-sample and zero-sample faults. Attached Figure Description

[0066] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0067] In the attached diagram:

[0068] Figure 1 This is a flowchart of the knowledge graph-based device fault diagnosis method of the present invention;

[0069] Figure 2 This is a structural block diagram of the knowledge graph-based device fault diagnosis system of the present invention;

[0070] Figure 3 This is a structural block diagram of the two-layer topology migration framework of the present invention.

[0071] Figure 4 This is a schematic diagram of the correction process for the fault propagation link in this invention.

[0072] Figure 5 This is a flowchart of the transfer reasoning process for zero-sample unknown faults in this invention. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0074] Several technical terms involved in this invention will be explained to facilitate understanding by those skilled in the art.

[0075] Fault Knowledge Graph: A data structure that stores knowledge in the field of fault diagnosis in the form of a graph. The nodes of the graph represent entities such as equipment components, fault modes, fault phenomena, and monitoring parameters, while the edges represent the semantic relationships between the above entities, and the weight of the edges represents the strength of the association.

[0076] Operating condition drift: refers to the deviation of the steady-state operating point of equipment from the calibrated state due to factors such as load changes, environmental parameter fluctuations, and differences in raw material properties during equipment operation. Operating condition drift is not a fault of the equipment, but it will cause changes in monitoring signals, which can easily lead to misjudgment by fixed threshold diagnostic methods.

[0077] Lifetime degradation coefficient: used to characterize the performance degradation of equipment; the larger the value, the more severe the aging.

[0078] Graph Attention Network (GAT): A neural network model based on graph structure. By assigning different attention weights to adjacent nodes, it achieves unequal weight aggregation of information in the graph. In this invention, it is used to calculate the importance weights of each edge on the fault propagation path.

[0079] Two-layer topology transfer framework: refers to a model framework that constructs a similarity measurement system at both the component level and the mechanism level, and uses this as a basis to transfer the knowledge graph topology of known faults to unknown or sparse fault scenarios.

[0080] Minimum maintenance unit: refers to equipment components that cannot be further disassembled or are not permitted to be further disassembled on-site during equipment maintenance operations.

[0081] Fault propagation delay: refers to the average time it takes for a fault to propagate from an upstream entity to a downstream entity when the equipment is in a baseline operating and healthy state.

[0082] Propagation reachability probability: This is used to characterize the probability that a fault will actually propagate from an upstream entity to a downstream entity along a certain link under the current operating conditions and equipment status.

[0083] Causal direction consistency score: used to characterize the degree of agreement between the occurrence time of the candidate root cause of the fault and the actual fault phenomenon alarm time after the propagation delay is added. The higher the score, the more likely the entity is to be the original root cause of the fault.

[0084] Comprehensive transfer similarity: A comprehensive metric that combines mechanistic semantic similarity and topological similarity to determine the feasibility of knowledge transfer between different components / faults.

[0085] Please refer to Figure 1 As an embodiment of the present invention, a device fault diagnosis method based on knowledge graph is provided;

[0086] This method first uses the acquired multi-source data on equipment to construct an initial knowledge graph of equipment faults;

[0087] When constructing the initial knowledge graph, it is necessary to build an initial ontology library for the fault knowledge graph of the target device. The ontology library contains a schema layer and a data layer for constructing the graph. The schema layer includes the following entity types and relation types:

[0088] Entity types include equipment, components, failure modes, failure symptoms, monitoring parameters, maintenance actions, etc.

[0089] Relationship types include location of occurrence, cause, accompaniment, indication, and resolution;

[0090] The data layer is constructed from multi-source equipment data, such as equipment maintenance work order records, fault case libraries, equipment operation manuals, and labeled samples of historical sensor monitoring data. During data processing, natural language processing methods such as entity recognition, relation extraction, and attribute filling are used to transform unstructured or semi-structured multi-source data into structured triples that can be stored in a graph database, thus forming an ontology library.

[0091] Specifically, the diagnostic method of the present invention includes the following steps:

[0092] Step S101:

[0093] The real-time operating conditions of the target equipment are obtained, the steady-state characteristics of the real-time operating conditions are extracted, and the state adjustment parameters of the target equipment are calculated.

[0094] The state adjustment parameters include the drift coefficient and the attenuation coefficient.

[0095] For extracting steady-state features, features are extracted from the sensor data collected by the equipment in real time using a sliding time window method. Steady-state features are used to reflect the characteristic quantities of the equipment's continuous operating state, such as the effective value of vibration (RMS), mean value, average temperature, and fundamental current amplitude. Steady-state features are used to evaluate the overall deviation of the operating condition.

[0096] For real-time operating conditions of equipment, it can also include extracting transient impact features. Transient features are used to reflect the characteristic quantities of sudden and instantaneous impacts on the equipment, such as the kurtosis, peak factor, and impulse index of vibration signals. They can identify the impact response caused by early local damage. When the transient impact features exceed the preset impact threshold, a temporary amplification factor can be applied to the operating condition drift coefficient to respond to sudden operating condition disturbances and improve the sensitivity of diagnosis.

[0097] In one implementation, the operating condition drift coefficient is obtained through the following steps:

[0098] Obtain the steady-state operating condition feature vectors for the current time period and the historical baseline time period, calculate the deviation between the current feature mean vector and the historical baseline time period feature mean vector, and standardize using the feature covariance matrix of the historical baseline time period to obtain the operating condition drift coefficient. , represented as:

[0099]

[0100] in, This represents the mean vector of steady-state characteristics for the current time period. This represents the mean vector of steady-state characteristics for a historical baseline period. represents the inverse of the covariance matrix; T represents the transpose of a vector.

[0101] The calculation of the above-mentioned working condition drift coefficient is essentially the calculation of the Mahalanobis distance between two mean vectors. This method takes into account the correlation between various feature dimensions and is not affected by the feature dimensions, so it can accurately reflect the overall deviation of multi-dimensional working conditions.

[0102] The degradation coefficient is the life-cycle degradation coefficient of the equipment. It is calculated based on the degradation relationship between the equipment's cumulative operating time and theoretical lifespan, combined with degradation parameters calibrated from historical fault data, to characterize the degree of equipment aging. In this embodiment, the Sigmoid function form is used:

[0103] In the early stages of operation, the attenuation coefficient is close to 0, indicating that the equipment is in a healthy state;

[0104] Near the median lifetime value, the factor rises rapidly;

[0105] In the later stages of operation, the factor approaches 1, indicating that the equipment has aged severely;

[0106] In this embodiment, the attenuation coefficient Represented as:

[0107]

[0108] Where T represents the cumulative operating time of the equipment. Represents the degradation rate coefficient. This represents the median life expectancy. and All were determined by the maximum likelihood estimation of the failure time of similar historical equipment.

[0109] In addition, for equipment types with relatively uniform degradation patterns, the attenuation coefficient can also be determined using a linear cumulative degradation model.

[0110] Step S102:

[0111] A preset initial map is reconstructed using state adjustment parameters. The reconstruction includes correcting the fault determination threshold and adjusting the fault propagation path. The state adjustment parameters are associated with and stored in the reconstructed map to obtain a dynamic map with the state adjustment parameters. The adjustment of the fault propagation path includes intensity correction and time delay correction.

[0112] By extracting the steady-state characteristics of the target equipment's real-time operating conditions and calculating the state adjustment parameters including the drift coefficient and attenuation coefficient, the state adjustment parameters are used to dynamically correct the fault judgment threshold and reconstruct the fault propagation link of the preset initial spectrum. The state adjustment parameters are then associated with and stored with the reconstructed spectrum to form a dynamic spectrum. This enables the judgment threshold and propagation attributes of the fault diagnosis spectrum to adapt to changes in the equipment's operating conditions and aging degree.

[0113] In traditional static knowledge graphs, the fault determination threshold is fixed. However, in real-world scenarios, when operating conditions drift or equipment ages, the baseline level of the monitoring signal will change systematically. Therefore, this application introduces a dynamic threshold correction based on the drift coefficient and attenuation coefficient, which can effectively avoid misjudgment or omission of diagnostic results due to fixed thresholds.

[0114] Specifically, the method for dynamically correcting the fault determination threshold in this invention includes:

[0115] Obtain the preset static fault judgment benchmark threshold. ;

[0116] Based on the aforementioned operating condition drift coefficient Determine the working condition penalty factor , The preset drift upper limit cutoff value is determined empirically based on the maximum allowable drift degree of the target equipment within its safe operating range.

[0117] Based on life cycle decay coefficient Determine the aging correction factor , ;

[0118] Calculate the dynamic fault determination threshold Represented as:

[0119]

[0120] In the formula, The operating condition drift coefficient, Indicates the baseline threshold. This is the lifecycle decay coefficient. Indicates the robustness coefficient under operating conditions. The aging sensitivity coefficient, and satisfying the following conditions. Not lower than the preset physical lower limit threshold The value of k2 makes It is positive;

[0121] When the operating conditions drift ( (increase) The threshold increases accordingly because normal fluctuations in operating conditions cause an overall rise in the monitoring signal. Maintaining the original threshold will frequently trigger false alarms. Therefore, appropriately increasing the threshold can effectively suppress such alarms. When the equipment ages (…), (increase) The threshold decreases as the fault tolerance of aging equipment decreases, and even small abnormal signals may indicate serious failures. Therefore, lowering the threshold can improve the detection sensitivity of early signs of degradation.

[0122] Meanwhile, this invention sets a physical lower limit threshold. ,make sure To prevent excessively low thresholds from generating false alarms in noisy environments; among other things, The settings are based on the noise floor and minimum resolvable signal level of the monitoring parameters;

[0123] Furthermore, the fault propagation chain in the knowledge graph is the foundation of causal reasoning about faults, and it includes not only semantic association strength attributes but also spatiotemporal attributes such as propagation delay. Under different operating conditions and aging stages of equipment, the causal strength and propagation speed of fault propagation will change, and adjusting only a single weight cannot truly reflect the actual laws of fault propagation.

[0124] like Figure 4 As shown, the dynamic reconstruction of the fault propagation links in the initial map by the present invention includes intensity correction and time delay correction;

[0125] Specifically, the intensity correction includes:

[0126] Obtain the original causal association weights between any two related entities in the initial graph. Original causal association weights For the initial graph construction phase, a pre-set association strength value is assigned to each relation edge based on expert experience or historical fault statistics.

[0127] Using the operating condition disturbance sensitivity coefficient and the aging conduction sensitivity coefficient to weight After correction, adaptive causal association weights are obtained. , represented as: ;in, This represents the sensitivity coefficient to operating disturbances. Indicates the aging conduction sensitivity coefficient; weighting Used to characterize the correlation strength of causal relationships in fault propagation;

[0128] and The value of ensures that the adaptive causal association weight remains within a reasonable range during normal device operation, avoiding unreasonable extreme values.

[0129] in addition:

[0130] In the operating condition disturbance sensitivity coefficient: the greater the operating condition drift, the stronger the noise interference on the fault causal relationship, and the lower the correlation weight accordingly;

[0131] In the aging conduction sensitivity coefficient: the more severe the equipment aging, the higher the degree of component deterioration, the easier it is for the fault to be conducted along the link, and the corresponding correlation weight increases accordingly.

[0132] Latency correction includes:

[0133] Obtain the standard propagation delay t0 of the corresponding propagation link in the initial map. The delay t0 is the average time for a fault to propagate from the upstream entity to the downstream entity when the device is in a baseline operating condition and a healthy state. It is obtained by calibrating historical fault data.

[0134] Based on the drift coefficient and attenuation coefficient, t0 is corrected to obtain the real-time propagation delay t. dyn , represented as:

[0135]

[0136] Wherein, k3 is the operating condition delay coefficient. Increased operating condition load will accelerate the fault evolution speed and shorten the propagation delay.

[0137] k4 is the aging delay coefficient; equipment aging accelerates fault propagation, and t dyn Not lower than the preset minimum delay limit t min To avoid excessive delays that do not conform to the laws of physics;

[0138] The values ​​of k3 and k4 are chosen such that, within the normal operating range of the equipment... , The average value is always positive, and t min Ensure the physical rationality under extreme operating conditions.

[0139] Step S103:

[0140] The real-time alarms of the target device are mapped to the dynamic graph, and the fault association subgraph is retrieved in the dynamic graph. After the fault association subgraph is topologically optimized, entity features are constructed based on state adjustment parameters. The propagation path is calculated by fusing entity features using a graph attention network to generate an initial candidate fault root cause set.

[0141] When the equipment system generates a real-time fault alarm, the real-time fault alarm information is first mapped to the dynamic graph. The monitoring parameters that trigger the alarm or the directly observed fault phenomenon are used as the entry node. The local subgraph associated with it is searched in the graph. The search range is within K hop neighbors (e.g., K=3) to ensure the real-time performance of the reasoning.

[0142] Topology optimization of the fault association subgraph includes:

[0143] The real-time propagation reachability probability p of the corresponding link is calculated based on the association weight and the real-time propagation delay. ij , represented as:

[0144]

[0145] in, This is the delay attenuation coefficient. The longer the propagation delay, the lower the probability that the fault will actually propagate along the link.

[0146] Weakly associated links with a real-time reachability probability lower than a preset reachability threshold are temporarily removed from the faulty association subgraph to reduce noise interference and improve the inference efficiency and accuracy of subsequent graph attention networks. The preset reachability threshold is set based on experience and is usually in the range of 0.1 to 0.3, so that only strongly associated links are retained for subsequent GAT inference.

[0147] If the core propagation path is broken due to the removal of weak links, an alternative propagation path is searched within a 3-hop range. The cumulative reachability probability of the alternative path is calculated. The alternative path whose cumulative reachability probability meets the threshold is added to the fault association subgraph as a hidden propagation link to ensure the integrity of the link.

[0148] Subsequently, the graph attention network GAT is used to perform weighted calculations on the fault propagation paths in the subgraph;

[0149] In the weighted computation of the propagation path in a graph attention network, the attention weights between any two adjacent entity nodes include:

[0150] Constructing multidimensional diagnostic composite feature vectors for each entity node , represented as:

[0151]

[0152] in, For real-time alarm feature encoding mapped by a fully connected layer, if the node corresponds to the monitoring parameter, then This is the vector mapped by the fully connected layer to the current anomaly level of the monitoring parameter; if it is a non-parametric node, it is filled with a learnable global embedding vector.

[0153] The learnable embedding of the lifecycle decay coefficient of entity i enables the network to perceive the aging state of the corresponding component of the entity node.

[0154] The learnable embedding of the working condition drift coefficient of entity i enables the network to perceive the working condition offset state of the device where the entity node is located.

[0155] φ(·) and ψ(·) are both learnable embedding mapping functions. In implementation, a single-layer fully connected network can be used to map the corresponding scalar parameters to Emb(A). i Embedding spaces with the same dimensions;

[0156] This indicates vector concatenation;

[0157] For the directed edge of fault propagation from component entity i to fault phenomenon entity j within the fault association subgraph, calculate the initial attention score. , represented as:

[0158]

[0159] in, To share the linear transformation matrix, Let T be the trainable attention parameter vector, and let T denote the transpose of the vector. This represents vector concatenation, with LeakyReLU as the activation function; Let i be the feature vector of entity i. Let j be the feature vector of entity j;

[0160] For all outgoing fault propagation directed edges of entity node i from the same source component, the attention weights between corresponding adjacent entity nodes are obtained through Softmax normalized exponential mapping. :

[0161]

[0162] In the formula, Let k be the set of neighboring nodes of node i in the fault association subgraph; k represents the node index. Let i represent the attention score between node i and any of its neighboring nodes k;

[0163] Therefore, this invention utilizes the calculated attention weights to perform weighted aggregation of fault propagation paths in the subgraph. After multiple rounds of information transmission, it obtains the aggregated representation vectors of each component entity or fault mode entity. Based on the similarity between the aggregated representation vectors and historical fault mode representations, it selects the Top-N candidate fault root causes to form a candidate fault root cause set.

[0164] In actual industrial scenarios such as mining equipment, historical samples of some failure modes are extremely scarce, which is referred to as small samples in this implementation. There may even be unknown failures that have never occurred before, such as zero-sample unknown failures.

[0165] Therefore, the diagnostic method of the present invention further includes:

[0166] Step S104:

[0167] A pre-constructed topology transfer model including components and mechanisms is adopted. When there are insufficient target fault samples, the mechanism similarity and topology similarity are corrected by state adjustment parameters. The corrected similarity is used to establish a cross-entity mapping relationship of the graph, and the candidate fault root cause set is expanded through transfer reasoning.

[0168] like Figure 3 As shown, the two-layer topology migration framework is a pre-built offline basic model. During the system deployment phase, the component topology layer, fault mechanism layer, and cross-layer mapping edge are built. During the online diagnosis phase, the model is directly called to perform migration inference.

[0169] By constructing a two-layer topology transfer framework that includes components and mechanisms, it is possible to handle different types of insufficient samples. The construction methods of the two-layer topology transfer framework include:

[0170] Construct a component topology layer, which uses the physical disassembly structure of the equipment as its skeleton, with nodes representing component entities and edges representing the assembly connection relationships and power transmission relationships between components;

[0171] A fault mechanism layer is constructed, which uses the physical mechanism of fault as the skeleton, with nodes representing fault mechanism entities and edges representing the causal evolution relationship between mechanisms;

[0172] A cross-layer mapping edge is established between the component topology layer and the fault mechanism layer, and the cross-layer mapping edge connects the component entity and the fault mechanism entity that the component may fail to.

[0173] During online diagnosis, state adjustment parameters are used to correct the mechanism similarity and topological similarity. Based on the corrected similarity, a cross-entity mapping relationship of the graph is established, and transfer reasoning is performed for small sample faults and zero sample faults respectively to expand the candidate fault root cause set.

[0174] For small-sample failure scenarios, when the number of historical samples for a certain type of failure is lower than a preset threshold (e.g., less than 5 cases), it is determined to be a small-sample failure scenario. In this case, the failure entity node already exists in the graph, but due to the sparse sample size, its connection relationships with other entities and attribute confidence are low. In this situation, the steps to expand the inference chain are executed, including:

[0175] Extract the semantic vector representation of the physical mechanism description text corresponding to the candidate fault root cause, and calculate the directional consistency metric between the semantic vector and the mechanism vector corresponding to the known fault entity, as the mechanism similarity. , represented as:

[0176]

[0177] In the formula, Represents the candidate fault mechanism vector; Given a known fault mechanism vector;

[0178] The source of the mechanism description text can be the failure mechanism field in the FMEA report or the fault cause description in the maintenance manual, and the text semantic vector can be obtained through a pre-trained BERT model or Word2Vec model.

[0179] Obtain the set of local adjacent nodes of the candidate faulty node. The set of locally adjacent nodes to the known faulty node Calculate the degree of two nodes. and The topological similarity is obtained by weighted fusion of Jaccard similarity with Laplacian smoothing and exponential decay similarity based on degree difference. Represented as:

[0180]

[0181] In the formula, This represents the smoothing constant (e.g., a value of 0.01). The parameter representing the degree difference scale. The degree of the candidate faulty node. Given the degree of the faulty node, , For preset weighting coefficients and ;

[0182] Furthermore, the mechanistic similarity and topological similarity are corrected using state adjustment parameters, specifically as follows:

[0183] Migration fusion weights are determined based on state conditioning parameters. , represented as:

[0184]

[0185] in, As the baseline fusion weight, The drift influence coefficient is... This is the attenuation effect coefficient;

[0186] The migration fusion weight increases with the increase of the drift coefficient and decreases with the increase of the attenuation coefficient;

[0187] The mechanistic similarity and topological similarity are weighted and fused using the aforementioned transfer fusion weights to obtain a comprehensive transfer similarity. , represented as:

[0188]

[0189] Known fault nodes whose comprehensive migration similarity exceeds a preset threshold are selected as migration sources. The missing connections in their neighborhoods are migrated to small sample fault nodes to complete the fault propagation chain and expand the candidate root cause set.

[0190] In addition, when the online monitoring signal shows an abnormal pattern, but no known fault entity can be matched in the existing knowledge graph, it is determined to be a zero-sample unknown fault situation; when the target fault is determined to be a zero-sample unknown fault, the monitoring signal corresponding to the current fault alarm is extracted, the monitoring signal is input into the pre-trained fault mechanism inference model, and the candidate physical failure mechanism description text and the corresponding confidence level are output.

[0191] The candidate mechanism description with the highest confidence level is selected to construct a virtual node of the failure mode in the dynamic graph;

[0192] Based on mechanism similarity matching, the known component entities that are closest to the virtual nodes are identified;

[0193] Based on the topological similarity matching, the known fault phenomenon entity that is closest to the virtual node is identified;

[0194] Then, cross-level connections are established between virtual nodes and matching entities to complete the fault reasoning link;

[0195] In one implementation, such as Figure 5 As shown, in the case of zero-sample unknown faults, the multi-dimensional monitoring signal corresponding to the current fault alarm is extracted. The multi-dimensional monitoring signal includes vibration spectrum envelope, temperature change slope, current harmonic distortion rate, etc., and is further input into the pre-trained fault mechanism inference model. It should be noted that the inference model adopts a sequence-to-sequence Transformer architecture and is trained by the multi-dimensional signal feature sequence and mechanism description text from historical fault cases.

[0196] In the specific training process, the inference model adopts a two-stage strategy of cross-device type transfer pre-training and target device fine-tuning:

[0197] The first stage is the transfer pre-training stage, which includes: collecting historical failure cases covering various types of electromechanical equipment, extracting paired samples containing clear mechanism descriptions and corresponding time period monitoring signals, and thus constructing a basic training set for cross-equipment failure mechanisms; using the basic training set to pre-train the Transformer architecture model, enabling the model to learn the mapping ability from monitoring signal feature sequences to failure mechanism description text.

[0198] The second stage is the fine-tuning stage. In this stage, a small number of labeled fault cases of the target diagnostic equipment are used as the fine-tuning set of the target equipment. Based on the pre-trained model, the parameters are fine-tuned to make the model adapt to the signal feature distribution and fault mechanism expression mode of the target equipment. A small learning rate is used in the fine-tuning process to reduce forgetting.

[0199] As an example, if a set of vibration spectrum envelope signals containing a specific modulation mode is input, the pre-trained inference model can output candidate mechanism descriptions and confidence scores for bearing inner ring fatigue spalling, periodic impact, and vibration amplitude modulation; after fine-tuning, the model can further distinguish the differences in mechanism expression under the specific bearing type and operating conditions of the target equipment.

[0200] Finally, the inference model outputs a set of candidate physical failure mechanism descriptions and corresponding confidence scores; for example, given a set of vibration spectra of a specific mode, the model may output two candidate mechanisms: periodic impact caused by inner ring failure and high-frequency friction caused by poor lubrication, and give confidence scores for each.

[0201] The candidate mechanism description with the highest confidence is selected, and a virtual node of the fault mode is constructed in the dynamic adaptive fault knowledge graph. The mechanism description is stored as a node attribute.

[0202] Based on mechanism similarity matching, the known component entities that are most similar to the mechanism vector of the virtual node are matched. Based on topological similarity matching, the known fault phenomenon entities that are most similar to the connection pattern of the virtual node are matched. Cross-level topological connection edges are established between the virtual node and the matched entities to complete the fault reasoning link and expand the candidate fault root cause set.

[0203] During expansion, virtual nodes and their connecting edges are incorporated into the fault association subgraph, enabling the graph attention network to infer the propagation path of the unknown fault and add it to the candidate fault root cause set. Preferably, after the subsequent maintenance personnel confirm the fault on-site, the verified virtual nodes can be converted into permanent nodes, which can realize the updating of the knowledge graph.

[0204] Step S105:

[0205] Sort the candidate fault root cause set, and determine and output the fault cause at the smallest maintenance unit level based on the sorting result;

[0206] Specifically, this invention obtains the aggregated propagation path score of the entity output by the graph attention network after multiple rounds of message passing for each candidate fault root cause entity in the root cause set.

[0207] For each candidate root cause entity, its potential fault occurrence time estimate is added to the corresponding real-time propagation delay to obtain the estimated phenomenon occurrence time; the estimated phenomenon occurrence time is compared with the first alarm timestamp of each fault phenomenon node to calculate the causal direction consistency score of the entity; if the estimated phenomenon occurrence time is systematically earlier than the actual alarm time, it indicates that the entity is consistent in the causal direction as the root cause of the fault.

[0208] In addition, the estimated time of occurrence of potential failure is derived by back-calculating the degradation rate represented by the attenuation coefficient of the corresponding component of the entity, combined with the degree of anomaly of the current alarm.

[0209] Specifically, the current degradation rate of the component is obtained based on the degradation curve corresponding to the life cycle decay coefficient. The degree of deviation of the current alarm signal from the benchmark threshold is used to characterize the severity of the fault. The onset time of the fault is calculated in reverse based on the degradation rate and the severity of the fault.

[0210] This invention introduces a causal direction consistency score based on alarm timestamps and real-time propagation delay. This is to distinguish between primary and secondary faults, and to score causal consistency. Represented as:

[0211]

[0212] In the formula, Let i represent the set of fault phenomenon nodes that can be reached in the fault association subgraph from the candidate fault root cause entity i. This is the timestamp of the first alarm at node j, where the fault occurs. This represents the estimated potential failure time associated with entity i; This represents the real-time propagation delay from entity i to faulty node j; For symbolic functions, This represents the attention weights output by the graph attention network;

[0213] If the occurrence time of a candidate fault entity, after adding the propagation delay, matches the actual alarm time of the fault phenomenon more closely, then its credibility as the primary fault (root cause) is higher, and the closer the score is to 1, the stronger the causal consistency.

[0214] Furthermore, when sorting the candidate root cause set, a weighted sum is performed based on the aggregated propagation path score, mechanism similarity, topological similarity, and causal direction consistency score to obtain the comprehensive confidence score of the entity. Finally, the candidate root cause entity can be sorted from high to low according to the comprehensive confidence score, and the top-ranked candidate root cause entity can be identified as the target root cause and output.

[0215] like Figure 2 As shown, according to another embodiment of the present invention, a diagnostic system for implementing a knowledge graph-based device fault diagnosis method is provided, comprising the following modules:

[0216] The state perception module is used to acquire the real-time operating conditions of the target device, extract the steady-state characteristics of the real-time operating conditions, and calculate the state adjustment parameters of the target device.

[0217] The graph reconstruction module is used to reconstruct a preset initial graph using state adjustment parameters. The reconstruction includes correcting the fault determination threshold and adjusting the fault propagation path. The state adjustment parameters are associated with and stored in the reconstructed graph to obtain a dynamic graph with the state adjustment parameters. The adjustment of the fault propagation path includes intensity correction and time delay correction.

[0218] The path reasoning module is used to map the real-time alarms of the target device to the dynamic graph, retrieve the fault association subgraph in the dynamic graph, perform topology optimization on the fault association subgraph, construct entity features based on state adjustment parameters, use graph attention network to fuse entity features to calculate the propagation path, and generate an initial candidate fault root cause set.

[0219] The migration expansion module is used to adopt a pre-built topology migration model that includes components and mechanisms. When there are insufficient target fault samples, the mechanism similarity and topology similarity are corrected by state adjustment parameters. The corrected similarity is used to establish a cross-entity mapping relationship of the graph, and the candidate fault root cause set is expanded through migration reasoning.

[0220] The sorting output module is used to sort the candidate fault root cause set, determine and output the fault cause at the smallest maintenance unit level based on the sorting result.

[0221] The modules mentioned above communicate with each other via a data bus or message queue, and can be deployed on an edge computing gateway or cloud server using a microservice architecture.

[0222] Another embodiment of the present invention provides a computer device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements a knowledge graph-based device fault diagnosis method as described in the above embodiments. The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc.

[0223] The computer device may include, but is not limited to, a processor and a memory.

[0224] Another embodiment of the present invention provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the knowledge graph-based device fault diagnosis method described in the above embodiment.

[0225] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A knowledge graph-based equipment fault diagnosis method, characterized in that, include: Steady-state features are extracted based on the real-time operating conditions of the target equipment, and the state adjustment parameters of the target equipment are calculated. A preset initial map is reconstructed using state adjustment parameters. The reconstruction includes correcting the fault determination threshold and adjusting the fault propagation path. The state adjustment parameters are then associated with the reconstructed map to obtain a dynamic map incorporating the state adjustment parameters. The adjustment of the fault propagation path includes intensity correction and time delay correction. The real-time alarms of the target device are mapped to the dynamic graph, and the fault association subgraph is retrieved in the dynamic graph. After the fault association subgraph is topologically optimized, entity features are constructed based on state adjustment parameters. The propagation path is calculated by fusing entity features using a graph attention network to generate an initial candidate fault root cause set. Using a pre-built topology transfer model, when there are insufficient target fault samples, the mechanistic similarity and topology similarity of the transfer reasoning are corrected by state adjustment parameters. The corrected similarity is used to establish a cross-entity mapping relationship of the graph, and the candidate fault root cause set is expanded through transfer reasoning. The candidate fault root cause set is sorted, and the fault cause at the smallest maintenance unit level is determined and output based on the sorting result.

2. The knowledge graph-based device fault diagnosis method according to claim 1, characterized in that, The state adjustment parameters include the drift coefficient and the attenuation coefficient; in: The drift coefficient is used to characterize the degree to which the current operating condition deviates from the reference operating condition; The attenuation coefficient is used to characterize the aging degree of the target equipment.

3. The knowledge graph-based device fault diagnosis method according to claim 2, characterized in that, In the reconstruction of the initial map, the step of correcting the fault determination threshold includes: Obtain the static fault determination benchmark threshold ; Determining the working condition penalty factor based on the drift coefficient , This is the preset drift upper limit cutoff value. This is the drift coefficient; The attenuation coefficient is used as the aging correction coefficient. ; Calculate the dynamic fault determination threshold Represented as: In the formula, Indicates the robustness coefficient under operating conditions. The aging sensitivity coefficient, and satisfying the following conditions. Not lower than the preset physical lower limit threshold .

4. The knowledge graph-based device fault diagnosis method according to claim 2, characterized in that, In the reconstruction of the initial spectrum, the intensity correction includes: Obtain the original causal association weights between any two related entities in the initial graph. ; Using the operating condition disturbance sensitivity coefficient and the aging conduction sensitivity coefficient to weight After correction, adaptive causal association weights are obtained. , represented as: ;in, This represents the sensitivity coefficient to operating disturbances. Indicates the aging conduction sensitivity coefficient; weighting Used to characterize the correlation strength of causal relationships in fault propagation.

5. The knowledge graph-based equipment fault diagnosis method according to claim 4, characterized in that, In the reconstruction of the initial map, the time delay correction includes: Obtain the standard propagation delay t0 of the corresponding propagation link in the initial map; Based on the drift coefficient and attenuation coefficient, t0 is corrected to obtain the real-time propagation delay t. dyn , represented as: Where k3 is the operating condition delay coefficient, k4 is the aging delay coefficient, and t dyn Not lower than the preset minimum delay limit t min .

6. The knowledge graph-based device fault diagnosis method according to claim 5, characterized in that, The topology optimization of the fault association subgraph includes: According to weight With delay Calculate the real-time propagation reachability probability p of the corresponding link. ij , represented as: in, This is the time delay attenuation coefficient; The weakly associated links are temporarily removed from the faulty association subgraph, and these weakly associated links are reached through real-time propagation of the reachability probability p. ij Determined if the value is below a preset reachable threshold; If the propagation path is broken due to the removal of weak links, an alternative path is retrieved within a preset number of hops and added to the fault association subgraph as an implicit link.

7. The knowledge graph-based device fault diagnosis method according to claim 6, characterized in that, In the calculation of the propagation path, the attention weights between any adjacent entity nodes are determined in the following way: Construct the feature vectors of each entity node ; For the directed edge of fault propagation from component entity i to fault phenomenon entity j within the fault association subgraph, calculate the initial attention score. , represented as: Where W is the shared linear transformation matrix, Let T be the trainable attention parameter vector, and let T denote the transpose of the vector. This represents vector concatenation, with LeakyReLU as the activation function; Let i be the feature vector of entity i. Let j be the feature vector of entity j; For all outgoing fault propagation directed edges of the same component entity node i, the attention weights between corresponding adjacent entity nodes are obtained through Softmax normalized exponential mapping. , represented as: In the formula, Let k be the set of neighboring nodes of node i in the fault association subgraph; k represents the node index. Let represent the attention score between node i and any of its neighboring nodes k.

8. The device fault diagnosis method based on knowledge graphs according to claim 1, characterized in that, The topology transfer model is a two-layer topology transfer framework, and the construction method of the two-layer topology transfer framework includes: Construct a component topology layer, which uses the physical disassembly structure of the equipment as its skeleton, with nodes representing component entities and edges representing the assembly connection relationships and power transmission relationships between components; A fault mechanism layer is constructed, which uses the physical mechanism of fault as the skeleton, with nodes representing fault mechanism entities and edges representing the causal evolution relationship between mechanisms; A cross-layer mapping edge is established between the component topology layer and the fault mechanism layer, and the cross-layer mapping edge connects the component entity and the fault mechanism entity that the component may fail to. During online diagnosis, state adjustment parameters are used to correct the mechanism similarity and topological similarity. Based on the corrected similarity, a cross-entity mapping relationship of the graph is established, and transfer reasoning is performed for small sample faults and zero sample faults respectively to expand the candidate fault root cause set.

9. The equipment fault diagnosis method based on knowledge graphs according to claim 1, characterized in that, The process of sorting the candidate fault root cause set and determining and outputting the fault cause at the smallest maintenance unit level based on the sorting result includes: For each candidate root cause entity in the root cause set, obtain the aggregated propagation path score of the entity output by the graph attention network after multiple rounds of message passing; For each candidate root cause entity, its potential failure occurrence time is estimated by adding the corresponding real-time propagation delay to obtain the predicted occurrence time of the phenomenon; the predicted occurrence time of the phenomenon is compared with the first alarm timestamp of each fault phenomenon node to calculate the causal direction consistency score of the entity. The overall confidence level of the entity is obtained by weighted summation of the aggregated propagation path score, mechanism similarity, topological similarity, and causal direction consistency score. Sort the targets by overall confidence level from high to low, determine the root cause of the fault based on the sorting results, and output the results.

10. A diagnostic system for implementing the knowledge graph-based device fault diagnosis method as described in any one of claims 1 to 9, characterized in that, Includes the following modules: The state perception module is used to extract steady-state features based on the real-time operating conditions of the target device and to calculate the state adjustment parameters of the target device. The graph reconstruction module is used to reconstruct a preset initial graph using state adjustment parameters. The reconstruction includes correcting the fault determination threshold and adjusting the fault propagation link. The state adjustment parameters are associated with the reconstructed graph to obtain a dynamic graph with the state adjustment parameters. The adjustment of the fault propagation link includes intensity correction and time delay correction. The path reasoning module is used to map the real-time alarms of the target device to the dynamic graph, retrieve the fault association subgraph in the dynamic graph, perform topology optimization on the fault association subgraph, construct entity features based on state adjustment parameters, use graph attention network to fuse entity features to calculate the propagation path, and generate an initial candidate fault root cause set. The migration expansion module is used to use a pre-built topology migration model. When there are insufficient target fault samples, it uses state adjustment parameters to correct the mechanism similarity and topology similarity of migration inference, and uses the corrected similarity to establish a cross-entity mapping relationship of the graph, thereby expanding the candidate fault root cause set through migration inference. The sorting output module is used to sort the candidate fault root cause set, determine and output the fault cause at the smallest maintenance unit level based on the sorting result.

Images