A Dual-Perspective Industrial Fault Diagnosis Method and System Combining Knowledge Graphs

By combining knowledge graphs and graph neural networks, and using one-dimensional convolutional networks and gated recurrent units to process time-series data, the system generates time-series dynamic vectors of device components and optimizes the scoring function. This solves the problems of insufficient real-time performance and interpretability in existing fault diagnosis methods, and achieves accurate fault diagnosis and a clear representation of fault propagation mechanisms.

CN123113831APending Publication Date: 2026-11-17ZHENGZHOU UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611474536.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing fault diagnosis methods cannot integrate real-time monitoring data and cannot perform differentiated modeling of heterogeneous relationships in knowledge graphs, resulting in diagnostic results that lack interpretability and real-time performance.

Method used

By combining knowledge graphs and graph neural networks, time-series monitoring data is processed through one-dimensional convolutional networks and gated recurrent units to generate time-series dynamic vectors of device components. A supernetwork is used to generate projection matrices and gate weights for triples, update node embedding vectors, optimize the scoring function to train the model, and identify fault propagation paths.

Benefits of technology

It enables accurate and real-time fault diagnosis of industrial equipment, improves the reliability and interpretability of diagnosis, and can clearly represent the fault propagation mechanism.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123113831A_ABST
    Figure CN123113831A_ABST
Patent Text Reader

Abstract

This invention relates to fault diagnosis, and in particular to a dual-perspective industrial fault diagnosis method and system combining knowledge graphs. The method includes: fusing time-series monitoring data of industrial equipment with structured data containing equipment topology and fault mechanisms; constructing a knowledge graph with equipment components as entities and topology and fault associations as relationships; extracting time-series vectors using a one-dimensional convolutional network and gated recurrent units as initial features of entity nodes; planning a dynamic graph neural network; generating relation-specific projection matrices and gating weights for knowledge graph triples through a hypernetwork; adjusting the message passing strength between nodes; and obtaining an embedding vector of the coupled real-time state of the equipment through end-to-end training. During fault diagnosis, abnormal component entities are first identified, and with fault type entities as targets, a beam search is performed in the vector space based on a scoring function to obtain the optimal path. The fault result and fault propagation explanation are output, achieving accurate and interpretable industrial fault diagnosis.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application belongs to the field of diagnostics, and in particular relates to a dual-perspective industrial fault diagnosis method and system that combines knowledge graphs. Background Technology

[0002] Industrial equipment contains hundreds or even thousands of interconnected components and is equipped with numerous sensors to monitor its operating status in real time, generating massive amounts of time-series monitoring data. How to utilize this high-dimensional time-series data, combined with the equipment's inherent physical structure and expert knowledge, to achieve rapid and interpretable fault diagnosis has become a key challenge in ensuring industrial production safety, improving equipment reliability, and reducing operation and maintenance costs. Traditional fault diagnosis methods rely on mechanistic models or expert systems. The former is difficult to model and cannot adapt to complex operating conditions, while the latter suffers from difficulties in knowledge base construction, high maintenance costs, and poor generalization ability. While data-driven methods based solely on deep learning, such as recurrent neural networks or convolutional neural networks, perform well in extracting fault features from time-series data, their opaque model decision-making process fails to incorporate prior structured knowledge such as equipment topology and fault propagation paths, resulting in diagnostic results lacking interpretability.

[0003] Knowledge graphs can model complex structured knowledge such as industrial equipment components, failure modes, and the physical connections and causal relationships between equipment components and failure modes in the form of "entity-relationship-entity" triples. Based on this, combining graph neural networks for representation learning can automatically learn low-dimensional vector representations of entities and relations in the knowledge graph, thereby detecting deep structural and semantic information. In message passing and node updating processes, existing models mostly use shared parameters or simplified aggregation methods to handle different types of relations, neglecting the unique semantics and influence weights of different relations. This homogeneous processing leads to information loss and fails to accurately model the fault propagation mechanism under heterogeneous relations. Existing methods often separate static knowledge graph construction from dynamic time-series data processing. Once the graph's embedding vectors are trained, they tend to be fixed, failing to deeply couple with the real-time operating status of the equipment. This results in insufficient sensitivity of the diagnostic model to changing operating conditions, limiting the real-time performance and accuracy of the diagnosis. Therefore, there is an urgent need for a new fault diagnosis method that can integrate real-time monitoring data and perform differentiated modeling of heterogeneous relations in the knowledge graph. Summary of the Invention

[0004] This invention proposes a dual-perspective industrial fault diagnosis method combining knowledge graphs to address the problems of existing fault diagnosis methods being unable to integrate real-time monitoring data and unable to perform differentiated modeling of heterogeneous relationships in knowledge graphs. The method includes the following steps: Acquire time-series monitoring data of industrial equipment within a preset time window, as well as structured data containing equipment topology and fault mechanisms; construct a knowledge graph with equipment components as entities and topology and fault mechanisms as relationships; process the time-series monitoring data using a one-dimensional convolutional network and gated recurrent units, extract time-series dynamic vectors representing the operating status of each equipment component within the time window, and use the time-series dynamic vectors as the initial features of the corresponding entity nodes in the knowledge graph; When aggregating and updating node features, for each triple in the knowledge graph, the hypernetwork generates a projection matrix and gating weights for the triple; the embedding vector of the relation is transformed using the projection matrix, and the message strength passed from the head entity to the tail entity is adjusted using the gating weights to update the embedding vector of the node; the entire model is trained by optimizing the scoring function used to evaluate the probability of the triple being true, thereby obtaining the entity and relation embedding vector coupled with the real-time operating status of the device; By continuously monitoring the time-series dynamic vectors of each device component, components with abnormal states are identified; the identified abnormal device component entities are used as starting nodes, and predefined fault type entities are used as the target node set. In the embedded vector space, a bundle search is performed using the above-mentioned scoring function as the path evaluation standard to find the optimal path from the starting node to any target node; the endpoint entity of the optimal path is used as the fault diagnosis result, and the path is used as the explanation for fault propagation.

[0005] Optionally, the step of processing the time-series monitoring data using a one-dimensional convolutional network and a gated recurrent unit to extract a time-series dynamic vector representing the operating state of each device component within the time window, and using the time-series dynamic vector as the initial feature of the corresponding entity node in the knowledge graph, includes: Represent the time-series monitoring data of length T and feature dimension F within the time window as a matrix; The matrix is ​​input into a one-dimensional convolutional layer with 64 kernels and a kernel size of 3, and a modified linear unit activation function is applied. The output of the one-dimensional convolutional layer is passed to a gated recurrent unit layer with 128 hidden units; The hidden state output of the gated loop unit at the last time step is taken as a temporal dynamic vector with a dimension of 128.

[0006] Optionally, the step of generating projection matrices and gate weights for triples using a hypernetwork includes: The d-dimensional embedding vectors of the head entity, relation entity, and tail entity are concatenated along the feature dimension to obtain a 3d-dimensional input vector. The 3D input vector is input into a supernetwork consisting of a two-layer fully connected neural network, wherein the hidden layer has a dimension of 4D and the output layer is divided into two branches; The first branch outputs a d×d vector and reshapes the vector into a d-row, d-column projection matrix; The second branch outputs a scalar value, which is then mapped to the 0-1 interval using the Sigmoid activation function as a gating weight.

[0007] Optionally, the transformation of the embedding vector of the relation using the projection matrix includes: When processing triples, the d-row, d-column projection matrix of the triples is obtained from the hypernetwork; The transformed relation vector is obtained by performing a matrix-vector multiplication operation between the d-dimensional embedding vector of the relation and the projection matrix.

[0008] Optionally, adjusting the message strength transmitted from the head entity to the tail entity using the gating weights and updating the node's embedding vector includes: By applying a preset message function to the head entity embedding vector and the transformed relation vector, the message vector from the head entity through the relation to the tail entity is calculated. Obtain the gating weights for the triples from the hypernetwork; The message vector is multiplied by the gating weights by a scalar to obtain the weighted message vector. When aggregating neighbor node information, the weighted message vector is used to update the node embedding vector of the tail entity.

[0009] Optionally, the step of training the entire model by optimizing the scoring function used to evaluate the probability of triples being true, thereby obtaining entity and relation embedding vectors coupled with the real-time operating state of the device, includes: The TransE model is used in conjunction with an exponential function as the scoring function. The score is the negative exponent of the norm of the difference vector obtained by subtracting the tail entity embedding vector from the sum of the head entity embedding vector and the transformed relation embedding vector. The cross-entropy loss function with negative sampling is used as the optimization objective. For each positive triplet, a preset number of negative triplets are constructed by randomly replacing the head entity or the tail entity. Use the Adam optimizer to train network parameters.

[0010] Optionally, the step of using the identified abnormal device component entities as starting nodes and the predefined fault type entities as the target node set, and performing a beam search in the embedding vector space using the aforementioned scoring function as the path evaluation criterion to find the optimal path from the starting node to any target node includes: Set the beam width to a preset integer; Starting from the starting node corresponding to the abnormal device component, in the first step, the score of all possible triples starting from the node is calculated, and the highest-scoring triples are retained to obtain the initial path; In each subsequent iteration, the endpoint entities of all current paths are expanded, the total score of all newly generated paths is calculated, the total score is the product of the scores of each triplet on the path, or equivalently the sum of the logarithmic values ​​of the scores of each triplet, and the highest total score is selected from all expanded paths as new candidate paths. Repeat the aforementioned iterative steps until the path length reaches the preset maximum depth, or the path endpoint reaches any entity in the target fault type entity set.

[0011] Furthermore, this invention also relates to a dual-perspective industrial fault diagnosis system incorporating knowledge graphs, comprising the following modules: An extraction module is used to acquire time-series monitoring data of industrial equipment within a preset time window, as well as structured data containing equipment topology and fault mechanisms; construct a knowledge graph with equipment components as entities and topology and fault mechanisms as relationships; and process the time-series monitoring data using a one-dimensional convolutional network and gated recurrent units to extract time-series dynamic vectors representing the operating status of each equipment component within the time window, and use the time-series dynamic vectors as the initial features of the corresponding entity nodes in the knowledge graph. The update module is used to generate a projection matrix and gating weights for each triple in the knowledge graph when aggregating and updating node features. The hypernetwork generates a projection matrix for the triple and gating weights for the triple. The embedding vector of the relation is transformed using the projection matrix and the message strength passed from the head entity to the tail entity is adjusted using the gating weights to update the embedding vector of the node. The entire model is trained by optimizing the scoring function used to evaluate the probability of the triple being true, thereby obtaining the entity and relation embedding vector coupled with the real-time operating status of the device. The execution module is used to identify components with abnormal states by continuously monitoring the time-series dynamic vectors of each device component; taking the identified abnormal device component entity as the starting node and the predefined fault type entity as the target node set, it performs a bundle search in the embedded vector space using the above scoring function as the path evaluation standard to find the optimal path from the starting node to any target node; taking the endpoint entity of the optimal path as the fault diagnosis result, and taking the path as the explanation of fault propagation.

[0012] Preferably, the step of processing the time-series monitoring data using a one-dimensional convolutional network and a gated recurrent unit to extract a time-series dynamic vector representing the operating state of each device component within the time window, and using the time-series dynamic vector as the initial feature of the corresponding entity node in the knowledge graph, includes: Represent the time-series monitoring data of length T and feature dimension F within the time window as a matrix; The matrix is ​​input into a one-dimensional convolutional layer with 64 kernels and a kernel size of 3, and a modified linear unit activation function is applied. The output of the one-dimensional convolutional layer is passed to a gated recurrent unit layer with 128 hidden units; The hidden state output of the gated loop unit at the last time step is taken as a temporal dynamic vector with a dimension of 128.

[0013] Preferably, the step of generating the projection matrix and gating weights for the triples by the hypernetwork includes: The d-dimensional embedding vectors of the head entity, relation entity, and tail entity are concatenated along the feature dimension to obtain a 3d-dimensional input vector. The 3D input vector is input into a supernetwork consisting of a two-layer fully connected neural network, wherein the hidden layer has a dimension of 4D and the output layer is divided into two branches; The first branch outputs a d×d vector and reshapes the vector into a d-row, d-column projection matrix; The second branch outputs a scalar value, which is then mapped to the 0-1 interval using the Sigmoid activation function as a gating weight.

[0014] Preferably, the transformation of the embedding vector of the relation using the projection matrix includes: When processing triples, the d-row, d-column projection matrix of the triples is obtained from the hypernetwork; The transformed relation vector is obtained by performing a matrix-vector multiplication operation between the d-dimensional embedding vector of the relation and the projection matrix.

[0015] Preferably, adjusting the message strength transmitted from the head entity to the tail entity using the gating weights and updating the node's embedding vector includes: By applying a preset message function to the head entity embedding vector and the transformed relation vector, the message vector from the head entity through the relation to the tail entity is calculated. Obtain the gating weights for the triples from the hypernetwork; The message vector is multiplied by the gating weights by a scalar to obtain the weighted message vector. When aggregating neighbor node information, the weighted message vector is used to update the node embedding vector of the tail entity.

[0016] Preferably, the step of training the entire model by optimizing the scoring function used to evaluate the probability of triples being true, thereby obtaining entity and relation embedding vectors coupled with the real-time operating state of the device, includes: The TransE model is used in conjunction with an exponential function as the scoring function. The score is the negative exponent of the norm of the difference vector obtained by subtracting the tail entity embedding vector from the sum of the head entity embedding vector and the transformed relation embedding vector. The cross-entropy loss function with negative sampling is used as the optimization objective. For each positive triplet, a preset number of negative triplets are constructed by randomly replacing the head entity or the tail entity. Use the Adam optimizer to train network parameters.

[0017] Preferably, the step of using the identified abnormal device component entities as the starting node and the predefined fault type entities as the target node set, and performing a beam search in the embedding vector space using the aforementioned scoring function as the path evaluation criterion to find the optimal path from the starting node to any target node includes: Set the beam width to a preset integer; Starting from the starting node corresponding to the abnormal device component, in the first step, the score of all possible triples starting from the node is calculated, and the highest-scoring triples are retained to obtain the initial path; In each subsequent iteration, the endpoint entities of all current paths are expanded, the total score of all newly generated paths is calculated, the total score is the product of the scores of each triplet on the path, or equivalently the sum of the logarithmic values ​​of the scores of each triplet, and the highest total score is selected from all expanded paths as new candidate paths. Repeat the aforementioned iterative steps until the path length reaches the preset maximum depth, or the path endpoint reaches any entity in the target fault type entity set.

[0018] This invention integrates temporal features representing the real-time operating status of equipment into knowledge graph entity nodes, achieving accurate and real-time representation of the industrial system's state. The planned representation learning model can generate relation-specific projections and weights for specific triples, thereby deeply exploring and identifying complex and diverse interactions between entities, improving the expressive power and accuracy of entity and relation embedding vectors. By searching for fault propagation paths in a high-quality embedding space, not only can the root cause of the fault be accurately located, improving the reliability of fault diagnosis, but the fault propagation mechanism can also be clearly represented, enhancing the interpretability of the diagnostic results. Attached Figure Description

[0019] Figure 1 A flowchart of a dual-perspective industrial fault diagnosis method that combines knowledge graphs; Figure 2 This is a schematic diagram of the network structure for temporal dynamic vector extraction. Figure 3 This is a schematic diagram of the hypernet structure and parameter generation. Detailed Implementation

[0020] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application 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 only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0021] It should be noted that, in this document, relational terms such as "first" and "second" are used merely 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 apparatus 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 apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0022] See in this application. Figure 1 As shown, a dual-perspective industrial fault diagnosis method combining knowledge graphs includes the following steps: S1, acquires the timing monitoring data of industrial equipment within a preset time window, as well as structured data containing equipment topology and fault mechanisms.

[0023] Sensor data, such as temperature, pressure, and vibration frequency, is acquired in real time from programmable logic controllers (PLCs) or distributed control systems (DCS) via the open platform communication unified architecture OPCUA protocol or message queue telemetry transmission (MQTT protocol). The data format is timestamp and value pairs, and it is stored in time-series databases such as InfluxDB or Prometheus. Structured data is parsed from equipment manuals, fault and maintenance records, and process flow diagrams (P and ID), and organized into CSV or JSON files containing head entities, relationships, and tail entities. The head and tail entities are equipment component IDs, and the relationships are physical connections, energy flow directions, or causal relationships.

[0024] S2, construct a knowledge graph with device components as entities and topology and fault mechanisms as relationships.

[0025] Using Python's NetworkX or RDFLib libraries, import triples from structured data files in batches to create directed graph data structures. Map entities and relation strings to unique integer IDs; store the constructed graph structure in a graph database such as Neo4j or Apache Jena for easy subsequent querying and access.

[0026] S3, and uses a one-dimensional convolutional network and a gated recurrent unit to process the time-series monitoring data, extracts the time-series dynamic vector representing the operating status of each device component within the time window, and uses the time-series dynamic vector as the initial feature of the corresponding entity node in the knowledge graph.

[0027] A sequence model is built using PyTorch or TensorFlow frameworks, with fixed-length time series segments as input. The model structure consists of: one or more one-dimensional convolutional layers using the ReLU activation function to extract local patterns; a max-pooling layer for dimensionality reduction and enhancing feature invariance; and one or more bidirectional gated recurrent unit (BiGRU) layers to detect long-term dependencies. The forward and backward hidden states of the last time step of the BiGRU are concatenated and mapped to a preset embedding dimension through a fully connected linear layer. (See reference...) Figure 2 As shown, the temporal dynamic vector is obtained.

[0028] As an optional embodiment, the step of processing the time-series monitoring data using a one-dimensional convolutional network and gated recurrent units to extract a time-series dynamic vector representing the operating state of each device component within the time window, and using the time-series dynamic vector as the initial feature of the corresponding entity node in the knowledge graph, includes: Represent the time-series monitoring data of length T and feature dimension F within the time window as a matrix; The matrix is ​​input into a one-dimensional convolutional layer with 64 kernels and a kernel size of 3, and a modified linear unit activation function is applied. The output of the one-dimensional convolutional layer is passed to a gated recurrent unit layer with 128 hidden units; The hidden state output of the gated loop unit at the last time step is taken as a temporal dynamic vector with a dimension of 128.

[0029] The combined model is a neural network containing a one-dimensional convolutional layer and a gated recurrent unit (ROU) layer. The network input is a two-dimensional matrix of size T×F, representing F-dimensional monitoring data of the device component within a time window T. The network structure is such that the input data passes through a one-dimensional convolutional layer, and the output sequence of that convolutional layer is fed into a gated recurrent unit (ROU) layer. The network output is the hidden state vector of the gated RNU layer at the last time step. The one-dimensional convolutional layer has 64 convolutional kernels, each with a size of 3, a stride of 1, and uses same padding and ReLU as the activation function. The subsequent gated RNU layer contains 128 hidden units. The 128-dimensional hidden state output vector of this gated RNU layer at the last time step is extracted as a temporal dynamic vector that comprehensively represents the operating state of the device component within this time window.

[0030] A one-dimensional convolutional network extracts local patterns and fundamental features from time-series data, such as brief peaks or oscillations in sensor readings. For example, when the time window T is 60 minutes (T = 60), and the monitored variables F are 10 (e.g., temperature, pressure, vibration, etc.), the input data is a 60×10 matrix. The convolutional layer can detect feature combinations within 3 minutes. The gated recurrent unit network then learns the changing trends and dynamic correlations over the entire 60-minute period based on these local features, such as a pattern of gradually increasing pressure accompanied by temperature anomalies. The output 128-dimensional vector encapsulates all the key information for the entire time period.

[0031] S4, when aggregating and updating node features, for each triple in the knowledge graph, the supernetwork generates a projection matrix and gating weights for the triple.

[0032] The hypernetwork is a two-layer feedforward neural network, i.e., a multilayer perceptron (MLP). To prevent overfitting, in one embodiment, the input is a concatenation of the head entity embedding vector and the tail entity embedding vector; the first hidden layer uses the LeakyReLU activation function, and the output layer does not use an activation function. The output of the hypernetwork is reshaped into two parts: a d×d projection matrix. , where d is the relation embedding dimension; and a scalar gating weight. .

[0033] As an optional embodiment, the generation of projection matrices and gated weights for triples by the hypernetwork includes: The d-dimensional embedding vectors of the head entity, relation entity, and tail entity are concatenated along the feature dimension to obtain a 3d-dimensional input vector. The 3D input vector is input into a supernetwork consisting of a two-layer fully connected neural network, wherein the hidden layer has a dimension of 4D and the output layer is divided into two branches; The first branch outputs a d×d vector and reshapes the vector into a d-row, d-column projection matrix; The second branch outputs a scalar value, which is then mapped to the 0-1 interval using the Sigmoid activation function as a gating weight.

[0034] The two-layer fully connected neural network is a supernetwork. The input to this network is a 3D concatenated vector, see [reference needed]. Figure 3 As shown, the network is formed by concatenating the d-dimensional embedding vector h of the head entity h, the d-dimensional embedding vector r of the relation r, and the d-dimensional embedding vector t of the tail entity t, denoted as [h;r;t]. The network structure consists of a hidden layer and an output layer with two independent branches. The output of the network includes a relation-specific projection matrix. and a gating weight The hidden layers of this network are 4d in dimension and use ReLU as the activation function. The first branch of the output layer outputs a... A 1-dimensional vector is reshaped into a d×d square matrix, which serves as the projection matrix. The second branch of the output layer outputs a scalar value, which is then passed through a sigmoid activation function and used as a gating weight. .

[0035] Assume the embedding dimension *d* of entities and relations is 128. The concatenated input vector has a dimension of 384. The hidden layer dimension of the hypernetwork is 512. The output vector of the first branch has a dimension of 16384, which, after reshaping, results in a 128×128 projection matrix. The second branch outputs a floating-point number between 0 and 1, such as 0.85, as the gating weight. This plan makes the projection matrix... and gating weights It is no longer a static attribute of relation r, but is jointly determined by the head entity h, relation r, and tail entity t that constitute the triple. This illustrates how the same relation acts in two different triples, pump A-pipe B and valve C-sensor D, depending on... Its manifestation and importance are derived from The manifestations are different, and the differences in the effects of different relationships on the same head and tail entity pairs can also be distinguished through the hypernetwork.

[0036] S5, the embedding vector of the relationship is transformed using the projection matrix, and the message strength transmitted from the head entity to the tail entity is adjusted using the gating weight, and the embedding vector of the node is updated.

[0037] Calculate the transformed relation vector , where r is the original embedding vector of the relation. Calculate the message passed from the head entity h to the tail entity t. ,in For a combinatorial function, such as the element-wise product in the DistMult scoring function. The new embedding vector of the tail entity t The update is performed by aggregating messages from all neighboring nodes. The aggregation function uses either summation or averaging. Where i represents all neighboring nodes of t. The activation function is such as Sigmoid or Tanh.

[0038] As an optional embodiment, the transformation of the embedding vector of the relation using the projection matrix includes: When processing triples, the d-row, d-column projection matrix of the triples is obtained from the hypernetwork; The transformed relation vector is obtained by performing a matrix-vector multiplication operation between the d-dimensional embedding vector of the relation and the projection matrix.

[0039] In the process of message passing and node updating in the graph neural network, when processing a specific triple h,r,t, the aforementioned hypernetwork module is invoked, and the embedding vectors of the head entity h and the tail entity t are input to obtain the d-row, d-column projection matrix generated for the specific triple. Simultaneously, a general, static d-dimensional embedding vector *r* of relation *r* is retrieved from the model's embedding layer. Matrix-vector multiplication is performed to compute the transformed relation vector. The newly generated relation vector The unique semantics of relation r in a specific context consisting of h and t were detected.

[0040] For example, suppose the embedding dimension d is 128. The general embedding vector r of the upstream relation is a 128-dimensional vector. When processing the triple pump A, upstream, tank B, the hypernetwork generates a 128×128 projection matrix based on the entity features of pump A and tank B. However, when dealing with another triplet valve C, upstream, and filter D, the hypernetic network generates a different 128×128 projection matrix. .pass and The calculations show that the model can distinguish between two different upstream relationships: pump-tank and valve-filter, making the transformed relationship vector... It can more accurately represent specific physical connections and functional dependencies, rather than a general concept.

[0041] As an optional embodiment, adjusting the message strength transmitted from the head entity to the tail entity using the gating weights and updating the node's embedding vector includes: By applying a preset message function to the head entity embedding vector and the transformed relation vector, the message vector from the head entity through the relation to the tail entity is calculated. Obtain the gating weights for the triples from the hypernetwork; The message vector is multiplied by the gating weights by a scalar to obtain the weighted message vector. When aggregating neighbor node information, the weighted message vector is used to update the node embedding vector of the tail entity.

[0042] Define and compute the message vector using a predefined message function, such as addition, to combine the embedding vector h of the head entity h with the relation vector transformed by the projection matrix. Adding them together yields the original message vector that was passed from the head entity h to the tail entity t via relation r. Obtain the gating weights by retrieving the gating weights (between 0 and 1) generated for the current triples h, r, t from the second branch of the supernetwork. Adjust the message strength by combining the original message vector m with the scalar gating weights. Perform multiplication to obtain the weighted message vector. During the aggregation phase of the graph neural network, the tail entity t collects all weighted messages from its neighboring nodes and uses these weighted messages to update its own embedding vector.

[0043] For example, in a triplet pressure sensor S, measuring a pipe P, if the real-time state of sensor S is reflected in the time-series dynamic vector and the subsequently updated entity embedding h, indicating that it is working normally, the supernetwork may generate a large gating weight. For example, a value of 0.9 allows measurement information to be smoothly transmitted to pipe P. If sensor S malfunctions, the embedding vector changes, and the hypernetwork may generate a gating weight close to 0, such as 0.1, to close or weaken the information path, preventing erroneous sensor data from contaminating the state representation of pipe P. This is crucial for isolating the effects of faults and improving the stability of the diagnostic model.

[0044] S6. The entire model is trained by optimizing the scoring function used to evaluate the probability of triples being true, thereby obtaining entity and relation embedding vectors coupled with the real-time operating status of the device.

[0045] Choose DistMult or ComplEx as the scoring function. For example, the score for DistMult is... Optimization is achieved using a negative sampling loss function. For each positive sample triplet, the head or tail entity is randomly replaced to generate several negative samples. The loss function is defined as follows: ,in The Sigmoid function is used; the Adam optimizer is employed to calculate and update the gradients of all parameters of the model using the backpropagation algorithm.

[0046] As an optional embodiment, the step of training the entire model by optimizing the scoring function used to evaluate the probability of triples being true, thereby obtaining entity and relation embedding vectors coupled with the real-time operating state of the device, includes: The TransE model is used in conjunction with an exponential function as the scoring function. The score is the negative exponent of the norm of the difference vector obtained by subtracting the tail entity embedding vector from the sum of the head entity embedding vector and the transformed relation embedding vector. The cross-entropy loss function with negative sampling is used as the optimization objective. For each positive triplet, a preset number of negative triplets are constructed by randomly replacing the head entity or the tail entity. Use the Adam optimizer to train network parameters.

[0047] Define a scoring function, adopting the idea of ​​the TransE model, combined with the transformed relation vector. The scoring function for the triple h,r,t is defined as follows: that is, vector The negative of the L2 norm (Euclidean distance) between t and t. A triplet should satisfy... At this point, the distance is small, and the score is close to 0, which is considered a high score. A loss function is constructed. A binary cross-entropy loss function with negative sampling is used. For each positive sample (true triplet) in the training set, K negative samples (non-existent triples) are constructed by randomly replacing the head or tail entity. The optimal value for K is 64 or 128. The goal of the loss function is to maximize the Sigmoid value of the positive sample score while minimizing the Sigmoid value of the negative sample score. Model optimization is performed. The Adam optimizer is used to train the entire model, including the temporal feature extraction module, embedding layer, hypernetwork, and all other parameters, using end-to-end gradient descent. The learning rate is set to 0.001, the batch size to 1024, and the weight decay can be set to... To prevent overfitting.

[0048] In each training iteration, the model extracts 1024 positive triples from the training set to form a batch. For each positive triple, such as pump A located in region 1, the model generates 64 negative samples, such as valve X located in region 1 or pump A located in region 5. The scores for one positive sample and 64 negative samples are calculated, and the total loss is calculated using the cross-entropy loss function. The backpropagation algorithm calculates the gradient based on this loss, and the Adam optimizer updates all learnable parameters of the model based on the gradient. Through tens of thousands of iterations, the model gradually learns embedding representations that result in high scores for positive samples and low scores for negative samples, enabling the entire embedding space to accurately represent the topology and fault mechanisms of industrial equipment.

[0049] S7 identifies components with abnormal states by continuously monitoring the timing dynamic vectors of each device component.

[0050] After model training is complete, the centroid or prototype vector of the temporal dynamic vector is calculated for each device component in normal state. During the monitoring phase, the Euclidean distance or cosine similarity between the newly generated temporal dynamic vector and the corresponding component prototype vector is calculated. When the distance exceeds a preset threshold or the similarity is lower than a preset threshold, the component is determined to be an abnormal component.

[0051] S8. Using the identified abnormal device component entities as the starting node and the predefined fault type entities as the target node set, a beam search is performed in the embedding vector space using the above scoring function as the path evaluation standard to find the optimal path from the starting node to any target node.

[0052] Set the bundle width B; starting from the starting node, explore all one-step reachable neighbor nodes in the knowledge graph, and compute h for each path. r Given the score s(h,r,t) for t, retain the B paths with the highest scores. In the next iteration, continue exploring from the endpoints of these B paths, calculating the total score of the new path. The total score is calculated by summing or multiplying the scores of all triples along the path. Again, select the B new path with the highest total score; repeat this process until the endpoint of the path is the target fault type entity or the maximum search depth is reached.

[0053] As an optional embodiment, the step of using the identified abnormal device component entities as starting nodes and the predefined fault type entities as the target node set, and performing a beam search in the embedding vector space using the aforementioned scoring function as the path evaluation criterion to find the optimal path from the starting node to any target node includes: Set the beam width to a preset integer; Starting from the starting node corresponding to the abnormal device component, in the first step, the score of all possible triples starting from the node is calculated, and the highest-scoring triples are retained to obtain the initial path; In each subsequent iteration, the endpoint entities of all current paths are expanded, the total score of all newly generated paths is calculated, the total score is the product of the scores of each triplet on the path, or equivalently the sum of the logarithmic values ​​of the scores of each triplet, and the highest total score is selected from all expanded paths as new candidate paths. Repeat the aforementioned iterative steps until the path length reaches the preset maximum depth, or the path endpoint reaches any entity in the target fault type entity set.

[0054] Parameter initialization: Set the bundle width to a positive integer B, for example, B=10, which represents the number of candidate paths retained in each search step. Set the maximum search depth. ,For example =5. Use the identified device component entity with an abnormal state as the starting node S. Path initialization, i.e., step 0. Create an initial path set containing only the starting node S itself, with a path score of 1, or equivalently, a logarithmic score of 0. Iterative expansion. From d=1 to... Perform a loop: For each of the B paths retained in the previous step, obtain the terminal node E, and search in the knowledge graph for all triples E,r whose head entity is E. The triples are extended to the end of the existing path to form a new candidate path of length d. Path evaluation and pruning: For all newly generated candidate paths, a total score is calculated. The total score is the product of the scores of all triples on the path calculated by the trained scoring function, or equivalently, the sum of the logarithmic values ​​of the scores of each triple. From all candidate paths, the B paths with the highest total score are selected as the retained paths for the current step. Termination condition judgment: After each iteration, check whether the terminal node of the retained path belongs to the predefined set of fault type entities. If the endpoint of a path is the target fault type, then the path is a potential diagnostic path. The search can terminate early or continue until the maximum depth is reached. The path that reaches the target and has the highest total score is selected as the diagnostic result.

[0055] For example, assuming the bundle width B=3, the starting node is the main pump A. Find all triples directly related to the main pump A, such as main pump A, drive, impeller; main pump A, connection, outlet valve; main pump A, monitoring, vibration sensor, and calculate the score for each triple. Assuming the three highest-scoring triples lead to the impeller, outlet valve, and lubrication system respectively, retain these three paths. Starting from the impeller, outlet valve, and lubrication system respectively, continue expanding the paths, potentially generating longer paths such as main pump A → impeller → wear, main pump A → outlet valve → jamming, etc. Recalculate the cumulative score for all new paths, and again retain only the top three. This process is repeated until the endpoint of a path is a pre-defined fault type entity such as bearing damage or cavitation failure. This highest-scoring complete path, such as main pump A → lubrication system → low oil pressure → bearing damage, not only provides the diagnostic result of bearing damage but also offers a clear explanation of the fault propagation chain.

[0056] S9, take the endpoint entity of the optimal path as the fault diagnosis result, and take the path as the explanation of fault propagation.

[0057] After the beam search is completed, the path with the highest total score is selected as the optimal path from all paths leading to the target node set. The last entity on this path, i.e., the specific fault type, is output as the diagnostic conclusion. The sequence of entities and relationships traversed by this path, such as pump A → pressure reduction → valve B → abnormal flow → system fault C, is provided to maintenance personnel as an interpretable fault propagation chain.

[0058] Based on any of the above embodiments, this application also provides a dual-view industrial fault diagnosis system incorporating knowledge graphs, comprising the following modules: An extraction module is used to acquire time-series monitoring data of industrial equipment within a preset time window, as well as structured data containing equipment topology and fault mechanisms; construct a knowledge graph with equipment components as entities and topology and fault mechanisms as relationships; and process the time-series monitoring data using a one-dimensional convolutional network and gated recurrent units to extract time-series dynamic vectors representing the operating status of each equipment component within the time window, and use the time-series dynamic vectors as the initial features of the corresponding entity nodes in the knowledge graph. The update module is used to generate a projection matrix and gating weights for each triple in the knowledge graph when aggregating and updating node features. The hypernetwork generates a projection matrix for the triple and gating weights for the triple. The embedding vector of the relation is transformed using the projection matrix and the message strength passed from the head entity to the tail entity is adjusted using the gating weights to update the embedding vector of the node. The entire model is trained by optimizing the scoring function used to evaluate the probability of the triple being true, thereby obtaining the entity and relation embedding vector coupled with the real-time operating status of the device. The execution module is used to identify components with abnormal states by continuously monitoring the time-series dynamic vectors of each device component; taking the identified abnormal device component entity as the starting node and the predefined fault type entity as the target node set, it performs a bundle search in the embedded vector space using the above scoring function as the path evaluation standard to find the optimal path from the starting node to any target node; taking the endpoint entity of the optimal path as the fault diagnosis result, and taking the path as the explanation of fault propagation.

[0059] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0060] The above description is merely a specific implementation of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A dual-perspective industrial fault diagnosis method combining knowledge graphs, characterized in that, Includes the following steps: Acquire time-series monitoring data of industrial equipment within a preset time window, as well as structured data containing equipment topology and fault mechanisms; construct a knowledge graph with equipment components as entities and topology and fault mechanisms as relationships; process the time-series monitoring data using a one-dimensional convolutional network and gated recurrent units, extract time-series dynamic vectors representing the operating status of each equipment component within the time window, and use the time-series dynamic vectors as the initial features of the corresponding entity nodes in the knowledge graph; When aggregating and updating node features, for each triple in the knowledge graph, the hypernetwork generates a projection matrix and gating weights for the triple; the embedding vector of the relation is transformed using the projection matrix, and the message strength passed from the head entity to the tail entity is adjusted using the gating weights to update the embedding vector of the node; the entire model is trained by optimizing the scoring function used to evaluate the probability of the triple being true, thereby obtaining the entity and relation embedding vector coupled with the real-time operating status of the device; By continuously monitoring the time-series dynamic vectors of each device component, components with abnormal states are identified; the identified abnormal device component entities are used as starting nodes, and predefined fault type entities are used as the target node set. In the embedded vector space, a bundle search is performed using the above-mentioned scoring function as the path evaluation standard to find the optimal path from the starting node to any target node; the endpoint entity of the optimal path is used as the fault diagnosis result, and the path is used as the explanation for fault propagation.

2. The method according to claim 1, characterized in that, The process of using a one-dimensional convolutional network and gated recurrent units to process the time-series monitoring data, extracting a time-series dynamic vector representing the operating state of each device component within the time window, and using the time-series dynamic vector as the initial feature of the corresponding entity node in the knowledge graph includes: Represent the time-series monitoring data of length T and feature dimension F within the time window as a matrix; The matrix is ​​input into a one-dimensional convolutional layer with 64 kernels and a kernel size of 3, and a modified linear unit activation function is applied. The output of the one-dimensional convolutional layer is passed to a gated recurrent unit layer with 128 hidden units; The hidden state output of the gated loop unit at the last time step is taken as a temporal dynamic vector with a dimension of 128.

3. The method according to claim 1, characterized in that, The generation of projection matrices and gated weights for triples by the hypernetwork includes: The d-dimensional embedding vectors of the head entity, relation entity, and tail entity are concatenated along the feature dimension to obtain a 3d-dimensional input vector. The 3D input vector is input into a supernetwork consisting of a two-layer fully connected neural network, wherein the hidden layer has a dimension of 4D and the output layer is divided into two branches; The first branch outputs a d×d vector and reshapes the vector into a d-row, d-column projection matrix; The second branch outputs a scalar value, which is then mapped to the 0-1 interval using the Sigmoid activation function as a gating weight.

4. The method according to claim 1, characterized in that, The transformation of the embedding vector of the relation using the projection matrix includes: When processing triples, the d-row, d-column projection matrix of the triples is obtained from the hypernetwork; The transformed relation vector is obtained by performing a matrix-vector multiplication operation between the d-dimensional embedding vector of the relation and the projection matrix.

5. The method according to claim 1, characterized in that, The step of adjusting the message strength transmitted from the head entity to the tail entity using the gating weights and updating the node's embedding vector includes: By applying a preset message function to the head entity embedding vector and the transformed relation vector, the message vector from the head entity through the relation to the tail entity is calculated. Obtain the gating weights for the triples from the hypernetwork; The message vector is multiplied by the gating weights by a scalar to obtain the weighted message vector. When aggregating neighbor node information, the weighted message vector is used to update the node embedding vector of the tail entity.

6. The method according to claim 1, characterized in that, The process involves training the entire model by optimizing a scoring function used to assess the probability of triplet formation, thereby obtaining entity and relation embedding vectors coupled with the real-time operating state of the device, including: The TransE model is used in conjunction with an exponential function as the scoring function. The score is the negative exponent of the norm of the difference vector obtained by subtracting the tail entity embedding vector from the sum of the head entity embedding vector and the transformed relation embedding vector. The cross-entropy loss function with negative sampling is used as the optimization objective. For each positive triplet, a preset number of negative triplets are constructed by randomly replacing the head entity or the tail entity. Use the Adam optimizer to train network parameters.

7. The method according to claim 1, characterized in that, The step of using the identified abnormal device component entities as starting nodes and the predefined fault type entities as the target node set, and performing a beam search in the embedding vector space using the aforementioned scoring function as the path evaluation criterion to find the optimal path from the starting node to any target node includes: Set the beam width to a preset integer; Starting from the starting node corresponding to the abnormal device component, in the first step, the score of all possible triples starting from the node is calculated, and the highest-scoring triples are retained to obtain the initial path; In each subsequent iteration, the endpoint entities of all current paths are expanded, the total score of all newly generated paths is calculated, the total score is the product of the scores of each triplet on the path, or equivalently the sum of the logarithmic values ​​of the scores of each triplet, and the highest total score is selected from all expanded paths as new candidate paths. Repeat the aforementioned iterative steps until the path length reaches the preset maximum depth, or the path endpoint reaches any entity in the target fault type entity set.

8. A dual-perspective industrial fault diagnosis system combining knowledge graphs, characterized in that, Includes the following modules: An extraction module is used to acquire time-series monitoring data of industrial equipment within a preset time window, as well as structured data containing equipment topology and fault mechanisms; construct a knowledge graph with equipment components as entities and topology and fault mechanisms as relationships; and process the time-series monitoring data using a one-dimensional convolutional network and gated recurrent units to extract time-series dynamic vectors representing the operating status of each equipment component within the time window, and use the time-series dynamic vectors as the initial features of the corresponding entity nodes in the knowledge graph. The update module is used to generate a projection matrix and gating weights for each triple in the knowledge graph when aggregating and updating node features. The hypernetwork generates a projection matrix for the triple and gating weights for the triple. The embedding vector of the relation is transformed using the projection matrix and the message strength passed from the head entity to the tail entity is adjusted using the gating weights to update the embedding vector of the node. The entire model is trained by optimizing the scoring function used to evaluate the probability of the triple being true, thereby obtaining the entity and relation embedding vector coupled with the real-time operating status of the device. The execution module is used to identify components with abnormal states by continuously monitoring the timing dynamic vectors of each device component; The identified abnormal device component entities are used as starting nodes, and the predefined fault type entities are used as the target node set. In the embedding vector space, a beam search is performed using the above scoring function as the path evaluation criterion to find the optimal path from the starting node to any target node. The endpoint entity of the optimal path is used as the fault diagnosis result, and the path is used as the explanation for fault propagation.

9. The system according to claim 8, characterized in that, The process of using a one-dimensional convolutional network and gated recurrent units to process the time-series monitoring data, extracting a time-series dynamic vector representing the operating state of each device component within the time window, and using the time-series dynamic vector as the initial feature of the corresponding entity node in the knowledge graph includes: Represent the time-series monitoring data of length T and feature dimension F within the time window as a matrix; The matrix is ​​input into a one-dimensional convolutional layer with 64 kernels and a kernel size of 3, and a modified linear unit activation function is applied. The output of the one-dimensional convolutional layer is passed to a gated recurrent unit layer with 128 hidden units; The hidden state output of the gated loop unit at the last time step is taken as a temporal dynamic vector with a dimension of 128.

10. The system according to claim 8, characterized in that, The generation of projection matrices and gated weights for triples by the hypernetwork includes: The d-dimensional embedding vectors of the head entity, relation entity, and tail entity are concatenated along the feature dimension to obtain a 3d-dimensional input vector. The 3D input vector is input into a supernetwork consisting of a two-layer fully connected neural network, wherein the hidden layer has a dimension of 4D and the output layer is divided into two branches; The first branch outputs a d×d vector and reshapes the vector into a d-row, d-column projection matrix; The second branch outputs a scalar value, which is then mapped to the 0-1 interval using the Sigmoid activation function as a gating weight.