A knowledge graph-based device fault prediction method and system

Abstract

The application discloses a kind of based on knowledge graph's equipment fault prediction method and system, it is related to equipment fault prediction technical field, including to equipment historical fault data and equipment topology and attribute data are preprocessed, entity node and relationship edge are defined in knowledge graph, form structured knowledge graph;The basic data in the process of equipment operation is preprocessed and extracts time-frequency feature and trend feature, input residual convolution network carries out feature extraction, and outputs basic feature vector, simultaneously using the residual convolution network to predict and calculate the basic residual of basic data, after the absolute value of the basic residual is normalized and handled, as sensor abnormal score.The method disclosed in the application fuses equipment topology structure, component attribute, historical fault experience and real-time sensor data into unified knowledge graph, realizes the integration modeling of semantic layer and data layer.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment failure prediction, and in particular to an equipment failure prediction method and system based on a knowledge graph. BACKGROUND

[0002] With the development of smart grids, industrial automation and smart operation, the stable operation of key industrial facilities such as power equipment, wind turbines and rail transit systems has high requirements for infrastructure safety. However, equipment is often affected by overload, aging, environmental disturbance and other factors during long-term operation, and is prone to potential failures such as overheating, overvoltage, abnormal vibration and component loosening. If early identification and accurate positioning of equipment failures cannot be achieved, it will easily cause accidents to spread, unplanned shutdowns, and even threats to personal safety.

[0003] In the prior art, commonly used equipment failure prediction methods include rule-based model-driven methods and time series-based machine learning methods. Rule-based model-driven methods such as expert systems and FMECA (Failure Mode and Effects Analysis) rely on manually set thresholds or rules and lack adaptability and generalization ability for new fault features.

[0004] Time series-based machine learning methods such as LSTM and AutoEncoder deep learning methods can model certain historical time series features, but often ignore structural information such as structural relevance between devices, component hierarchical topology and semantic dependence between sensors, limiting prediction accuracy and making the model difficult to interpret.

[0005] Therefore, how to integrate various data, build a graph model with semantic association, structural relationship and computational reasoning ability, and then achieve high robustness and high interpretability of equipment failure prediction has become a key technical problem that needs to be solved in the current field of industrial intelligent monitoring. SUMMARY

[0006] In view of the above existing problems, the present application is proposed.

[0007] Therefore, the present application provides an equipment failure prediction method based on a knowledge graph.

[0008] To solve the above technical problems, the present application provides the following technical solutions:

[0009] In a first aspect, the present application provides an equipment failure prediction method based on a knowledge graph, which comprises,

[0010] Step 1: Preprocess the device historical failure data and device topology and attribute data, define entity nodes and relationship edges in the knowledge graph, and form a structured knowledge graph;

[0011] Step two, pre-process the basic data of the device and extract features to obtain the basic feature vector; predict the basic data and calculate the basic residual, and obtain the sensor anomaly score based on the basic residual;

[0012] Step three, embed each entity node in the knowledge graph to obtain an entity vector, map the basic feature vector and the sensor anomaly score to the corresponding feature node, and calculate the cosine similarity and Euclidean distance between the feature node and each fault type node to form the joint mapping feature of each fault type;

[0013] Step four, construct a graph neural network based on the structure of the knowledge graph, input the joint mapping feature for message propagation, define a score for each fault type node, and map the score to obtain the probability of occurrence of each fault type;

[0014] Step five, input the sensor anomaly score and entity vector obtained by processing the real-time collected basic data in steps two and three into the graph neural network to generate the prediction probability of each fault type, and output a fault warning signal when the prediction probability of any fault type is greater than a preset threshold.

[0015] As a preferred scheme of the device fault prediction method based on the knowledge graph, wherein: the score is a linear combination of the sensor anomaly score and the updated weight between nodes after message propagation.

[0016] As a preferred scheme of the device fault prediction method based on the knowledge graph, wherein: the basic data includes voltage, current, temperature and vibration amplitude, the device historical fault data includes device fault repair records and alarm logs, and the device topology and attribute data includes the hierarchical structure, connection relationship and specification parameters of the device and its components.

[0017] Defining entity nodes and relationship edges in the knowledge graph means defining entity nodes representing devices, components, sensor measurement points and fault types, and relationship edges representing subordinate relationships, monitoring relationships and fault association relationships.

[0018] As a preferred scheme of the device fault prediction method based on the knowledge graph, wherein: the cosine similarity represents the vector direction correlation between the sensor feature and the fault type, and the Euclidean distance represents the spatial distance between the sensor feature and the fault type vector.

[0019] As a preferred scheme of the device fault prediction method based on the knowledge graph, wherein: the pre-processing of the device historical fault data includes using a natural language processing method to extract fault entities, fault types and occurrence timestamps in the device fault repair records and alarm logs to generate structured fault event data.

[0020] The preprocessing of the device topology and attribute data includes extracting and generating structured topology data representing the hierarchy, connection relationship and specification parameters of the device and its components;

[0021] The preprocessing of the basic data includes band-pass filtering and normalization processing of voltage, current and vibration amplitude data, sliding mean filtering and normalization processing of temperature data, and extracting trend features in the sliding window, the trend features including sliding window mean, standard deviation and trend slope.

[0022] As a preferred scheme of the device fault prediction method based on the knowledge graph, wherein: after forming the structured knowledge graph, the Granger causality analysis is used to determine the causal relationship strength between the sensor feature data and the fault event, and the knowledge graph is updated, including the following steps,

[0023] Collecting historical sensor measurement point data and fault events at the same time to form a time series data set, the fault event refers to whether a certain fault occurs;

[0024] For each type of sensor data and the fault event in the historical record, a Granger causal relationship model is established respectively, represented as,

[0025]

[0026] In the formula, refers to the probability of failure K at time point t, represents the intercept term, which is a constant, is the fault autoregressive coefficient, represents the influence coefficient of the i th sensor channel on the fault at lag time l, and L is the lag order, refers to the observation value of the i th sensor at time t-l, refers to the state of the fault type k at time t-l;

[0027] The best lag order L is dynamically determined by the Akaike information criterion, and the coefficient sequence of the regression result is jointly tested, the statistical quantity is calculated, the significance threshold is set, and if the significance is met, it is determined that the feature data xi has a significant causal relationship with the fault event;

[0028] The confirmed causal relationship is explicitly added to the knowledge graph.

[0029] As a preferred scheme of the device fault prediction method based on the knowledge graph, wherein: the joint mapping features of each fault type include,

[0030] ​By combining outlier scores, Euclidean distance, and cosine similarity, a joint weighted contribution is constructed, denoted as:

[0031]

[0032] In the formula, For sensor anomaly score, , These are the weighting coefficients. For cosine similarity, The distance is Euclidean.

[0033] For each fault type node By aggregating the contributions from all related sensors, their joint mapping features are obtained, denoted as:

[0034]

[0035] In the formula, m represents the number of sensors, meaning that the time-series characteristics of each fault node are derived from the embedding vectors of all its sensor nodes. Perform a weighted summation, with the weight representing the channel's abnormal contribution at the current moment. .

[0036] In a preferred embodiment of the knowledge graph-based device fault prediction method of the present invention, the score is represented as a linear combination of the sensor anomaly score and the weights updated by message propagation between nodes.

[0037]

[0038] In the formula, As weight, For sensor anomaly scores, The sensor anomaly score at time t;

[0039] The probability of each fault type occurring is expressed as follows:

[0040]

[0041] In the formula, To predict the type of failure that will occur at time t The probability is given by K, where K is the total number of fault types and j is the traversal parameter for the fault types.

[0042] In a preferred embodiment of the knowledge graph-based equipment fault prediction method of the present invention, after constructing the graph neural network, the weight parameters of the graph neural network are updated using the cross-entropy loss function based on historical equipment operation data and fault event labels.

[0043] Secondly, this invention provides a knowledge graph-based device fault prediction system, comprising,

[0044] The map construction module is used to collect historical fault data of equipment, as well as equipment topology and attribute data, and to perform structured processing on the collected data.

[0045] The feature extraction module is used to collect and process basic data during equipment operation.

[0046] The vector mapping module is used to generate node vector representations based on the structural relationships between nodes in the knowledge graph and to achieve feature fusion.

[0047] The fault reasoning module is used to build a graph neural network model based on the knowledge graph and to perform fault type prediction and reasoning.

[0048] The real-time early warning module is used to receive real-time collected basic data and perform fault prediction based on the trained graph neural network.

[0049] The beneficial effects of this invention are as follows: it integrates equipment topology, component attributes, historical fault experience, and real-time sensor data into a unified knowledge graph, achieving integrated modeling of the semantic and data layers. Through Granger causality analysis, it automatically mines prior causal relationships between sensor features and fault types from historical data and embeds them as graph edge relationships, improving the interpretability and accuracy of the prediction model. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a flowchart of the knowledge graph-based device fault prediction method in Example 1. Detailed Implementation

[0052] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0053] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0054] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0055] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a device fault prediction method based on knowledge graphs, including the following steps:

[0056] Step 1: Collect historical fault data and equipment topology and attribute data. After preprocessing the two types of data, define entity nodes and relationship edges in the knowledge graph. Store the entities and relationships in the graph database to form a structured knowledge graph.

[0057] Specifically, historical fault data of smart grid equipment is collected from the operation and maintenance system, including equipment fault repair records and alarm logs, forming raw data in text format. The Transformer-BERT natural language processing model is then used to perform text preprocessing and structured processing on the collected historical fault data. The specific steps are as follows:

[0058] Sentence segmentation and word segmentation are performed on the text data;

[0059] Named Entity Recognition (NER) technology is used to automatically identify and extract entity information such as "device name", "fault type", "fault time", "fault location" and "fault cause" from fault records;

[0060] Then, using a joint entity relationship extraction model (such as CasRel), the relationships between entities are extracted, such as the "occurrence" relationship between fault type and equipment component, and the "causal" relationship between fault type and fault cause.

[0061] The extracted data is then formed into a standard structured triplet form, such as (entity 1, relation, entity 2).

[0062] In a preferred embodiment, the topology and attribute data of the device specifically include the hierarchical structure, connection relationships and specification parameters of the device and its components, which are usually represented by structured tables or tree data.

[0063] The topology data is uniformly encoded and assigned a unique ID to each device and its components; the device attribute data (such as rated voltage, rated current, etc.) is uniformly normalized.

[0064] The processed topology and attribute data are used to form structured entity relationship data, such as (Transformer 002, including components, Fan 01) and (Transformer 002, parameters, rated voltage 220kV).

[0065] Furthermore, defining entity nodes and relation edges in a knowledge graph means defining entity nodes that represent devices, components, sensor measurement points, and fault types, as well as relation edges that represent subordinate relationships, monitoring relationships, and fault association relationships.

[0066] After forming a structured knowledge graph, Granger causality analysis is used to determine the strength of the causal association between sensor feature data and fault events, and the knowledge graph is updated, including the following steps.

[0067] Historical sensor measurement data and fault events at the same time are collected to form a time series data set, wherein the fault event refers to whether a certain fault has occurred.

[0068] For each type of sensor data and the fault events in historical records, a Granger causality model is established, represented as follows:

[0069]

[0070] In the formula, It refers to the probability that failure K will occur at time point t. The intercept term is a constant. The fault autoregressive coefficient, This represents the influence coefficient of the i-th sensor channel on the fault over the lag time l, where L is the lag order. This refers to the observation value of the i-th sensor at time tl. This refers to the state of fault type k at time tl;

[0071] The optimal lag order L is dynamically determined using the Akaike information criterion, denoted as:

[0072]

[0073] In the formula, k is the number of parameters, that is... and The total number of values, W is the maximum likelihood function value, and the combination of lag orders that minimizes AIC / BIC is selected from the range of L=1~5 to avoid overfitting or underfitting.

[0074] For the regression results The coefficient sequences are subjected to joint testing, and the statistical hypothesis is:

[0075]

[0076] At least one exists. .

[0077] Calculate the statistic.

[0078]

[0079] In the formula, To constrain the sum of squared residuals in the model, Let L be the sum of squared residuals of the unconstrained model, L be the lag order, n be the sample size, and k be the number of parameters.

[0080] A significance threshold is set; in this embodiment, the threshold is set to 0.05. If the p-value corresponding to the joint F-test is less than 0.05, the causal relationship is determined to be valid, and the feature data x is determined. i It has a significant causal relationship with the failure event;

[0081] The confirmed causal relationships are explicitly added to the knowledge graph.

[0082] By employing Granger causal analysis after forming a structured knowledge graph, statistical modeling of sensor historical data and fault events is carried out to quantify the intensity of their causal influence and update the corresponding edge weights in the graph. This significantly improves the semantic accuracy and quantifiable interpretability of the relationships in the knowledge graph, thereby enhancing the learning ability of subsequent graph neural network models to the fault occurrence mechanism, improving the accuracy and interpretability of fault prediction results, and realizing the dynamic coupling update of the graph and the operating status, thus improving the overall adaptive capability of the system.

[0083] Step 2: Collect basic data during equipment operation. After preprocessing the basic data, extract time-frequency features and trend features, input them into the ResNet-50 residual convolutional network for feature extraction, and output basic feature vectors. At the same time, use the ResNet-50 residual convolutional network to predict the basic data. Determine the basic residual by the difference between the actual measured value and the predicted value. Normalize the absolute value of the basic residual and use it as the sensor anomaly score.

[0084] Specifically, the basic data includes voltage, current, temperature, and vibration amplitude. Preprocessing of this basic data includes bandpass filtering of the voltage, current, and vibration amplitude data to remove low-frequency drift and high-frequency interference, retaining their fault response frequency band. A recommended filtering band is 5Hz~100Hz. Afterwards, normalization is performed, and a Short-Time Fourier Transform (STFT) is used to convert the one-dimensional time-series data into a two-dimensional time-spectrum tensor. A sliding window with a window length of 256 and an overlap length of 128 is used to divide the original signal, extracting and forming a two-dimensional tensor, which serves as the input features for ResNet.

[0085] Temperature data undergoes a moving average filter with a 60-second window to eliminate minor fluctuations and preserve the trend. Normalization is then applied, and a moving average statistical feature extraction method is used to extract trend features within the moving window. These trend features include the moving window mean, standard deviation, and trend slope, which are combined into a vector to form the feature representation of the temperature channel. It should be noted that the calculation methods for the moving window mean, standard deviation, and trend slope are existing technologies and will not be elaborated upon here.

[0086] Step 3: Embed each entity node in the knowledge graph using the Node2Vec algorithm to obtain entity vectors. Map the basic feature vectors and sensor anomaly scores to the corresponding feature nodes. Calculate the cosine similarity and Euclidean distance between the feature nodes and nodes of each fault type based on the entity vectors, thus forming a joint mapping feature for each fault type.

[0087] Specifically, embedding each entity node in the knowledge graph using the Node2Vec algorithm to obtain the entity vector involves the following steps:

[0088] Export the representations of all nodes and edges from the knowledge graph constructed in step one, and initialize the parameters, setting the following embedding parameters to their default optimal values:

[0089] Embedding dimension d=128;

[0090] Walking stride length l = 80;

[0091] The number of traversal paths sampled at each node is r=10;

[0092] The probability of returning is p=1.0, and the probability of forward exploration is q=0.5, which encourages the discovery of nodes with similar structures and functions.

[0093] After executing Node2Vec, the embedding matrix is ​​obtained:

[0094]

[0095] in, This represents the embedding vector of node v in the graph, where V is the set of nodes and E is the set of edges. It is the space of real numbers.

[0096] Furthermore, the cosine similarity represents the vector direction correlation between sensor features and fault types.

[0097]

[0098] In the formula, Let i be the feature vector of sensor i. For the faulty node Embedded vector.

[0099] The Euclidean distance represents the spatial distance between sensor features and the fault type vector.

[0100] .

[0101] Combining the outlier scores with the two indicators mentioned above, a joint weighted contribution is constructed, denoted as:

[0102]

[0103] In the formula, For sensor anomaly score, , For example, the weighting coefficients are... , .

[0104] For each fault type node By aggregating the contributions from all related sensors, their joint mapping features are obtained, denoted as:

[0105]

[0106] In the formula, m represents the number of sensors, meaning that the time-series characteristics of each fault node are derived from the embedding vectors of all its sensor nodes. Perform a weighted summation, with the weight representing the channel's abnormal contribution at the current moment. .

[0107] Step four: In this embodiment, a Graph Convolutional Network (GCN) is used as the basic model. A graph neural network is constructed based on the knowledge graph, and message propagation is performed using the joint mapping features as input. A score is defined for each fault type node. The score is a linear combination of the sensor anomaly score and the weights updated after message propagation between nodes, expressed as follows:

[0108]

[0109] In the formula, As weight, The sensor anomaly score is given.

[0110] The scores of all fault type nodes are combined into a score vector, which is then mapped to a probability distribution using the Softmax function to obtain the probability of each fault type occurring, denoted as:

[0111]

[0112] In the formula, To predict the type of failure that will occur at time t The probability of K is the total number of fault types.

[0113] Using historical equipment operation data and fault event labels, the weight parameters of the graph neural network are trained, optimized, and regularized using the cross-entropy loss function. Specifically, a training set is constructed from historical data, where each sample consists of joint features of nodes and a label of the actual fault type. The cross-entropy loss function is defined (this is existing technology and will not be elaborated here), and the Adam optimizer is used with an initial learning rate set to 0.001. The regularization term penalizes the weight size to prevent overfitting. The training process includes: inputting graph data in mini-batch mode; updating the weights in each iteration; monitoring the cross-entropy loss using a validation set; and stopping training after reaching a set number of iterations. It should be noted that this step is existing technology and will not be elaborated upon.

[0114] Step 5: After processing the real-time collected basic data in Steps 2 and 3, the sensor anomaly score and entity vector are obtained as inputs and fed into the graph neural network to generate the predicted probability of each fault type. When the predicted probability of any fault type is greater than a preset threshold, a fault warning signal is output.

[0115] In summary, the method described in this invention integrates equipment topology, component attributes, historical fault experience, and real-time sensor data into a unified knowledge graph, achieving integrated modeling of the semantic and data layers. Granger causality analysis automatically mines prior causal relationships between sensor features and fault types from historical data and embeds them as graph edge relationships, improving the interpretability and accuracy of the prediction model. Furthermore, a ResNet-50 deep network is introduced to extract basic feature vectors from the time-frequency and trend domains, combining node embedding and anomaly scoring to construct fused features, providing information-rich input to the graph neural network.

[0116] This embodiment also provides a knowledge graph-based device fault prediction system, including:

[0117] The knowledge graph construction module is used to collect historical fault data and equipment topology and attribute data, and to perform structured processing on the collected data. Specifically, it is used to collect historical fault data and equipment topology and attribute data, preprocess the two types of data, define entity nodes and relation edges in the knowledge graph, and store the entities and relations through a graph database to form a structured knowledge graph.

[0118] The feature extraction module is used to collect and process basic data during equipment operation. Specifically, it collects basic data during equipment operation, preprocesses the basic data to extract time-frequency features and trend features, inputs them into a ResNet-50 residual convolutional network for feature extraction, outputs a basic feature vector, and uses the ResNet-50 residual convolutional network to predict the basic data. The difference between the actual measured value and the predicted value is used to determine the basic residual. The absolute value of the basic residual is normalized and used as the sensor anomaly score.

[0119] The vector mapping module is used to generate node vector representations based on the structural relationships of nodes in the knowledge graph and to achieve feature fusion. Specifically, it is used to embed each entity node in the knowledge graph based on the Node2Vec algorithm to obtain entity vectors, map the basic feature vectors and sensor anomaly scores to the corresponding feature nodes, and calculate the cosine similarity and Euclidean distance between the feature nodes and nodes of each fault type based on the entity vectors to form joint mapping features for each fault type.

[0120] The fault reasoning module is used to construct a graph neural network model based on the knowledge graph and complete the prediction and reasoning of fault types. Specifically, it is used to construct a graph neural network based on the knowledge graph, use the joint mapping features as input for message propagation, define a score for each fault type node, the score is a linear combination of the sensor anomaly score and the weights updated by message propagation between nodes, map the score through the Softmax function to obtain the probability of each fault type occurring, and use historical equipment operation data and fault event labels to train, optimize and regularize the weight parameters of the graph neural network using the cross-entropy loss function.

[0121] The real-time early warning module is used to receive real-time collected basic data and perform fault prediction based on the trained graph neural network. Specifically, it is used to process the real-time collected basic data through steps two and three to obtain sensor anomaly scores and entity vectors as inputs, which are then input into the graph neural network to generate prediction probabilities for each fault type. When the prediction probability of any fault type is greater than a preset threshold, a fault early warning signal is output.

[0122] This embodiment also provides a computer device applicable to the device fault prediction method based on knowledge graphs, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the device fault prediction method based on knowledge graphs as proposed in the above embodiment.

[0123] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0124] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the knowledge graph-based device fault prediction method as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0125] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A knowledge graph-based method for predicting equipment failures, characterized in that: include, Step 1: Preprocess the historical fault data of the equipment and the topology and attribute data of the equipment, and define entity nodes and relationship edges in the knowledge graph to form a structured knowledge graph; Step 2: Preprocess the basic data of the equipment and extract features to obtain basic feature vectors; predict the basic data and calculate the basic residuals; obtain sensor anomaly scores based on the basic residuals. Step 3: Embed each entity node in the knowledge graph to obtain entity vectors. Map the basic feature vectors and sensor anomaly scores to the corresponding feature nodes. Calculate the cosine similarity and Euclidean distance between the feature nodes and nodes of each fault type to form the joint mapping features for each fault type. Step 4: Construct a graph neural network based on the knowledge graph, use the joint mapping features as input for message propagation, define a score for each fault type node, and map the scores to obtain the probability of each fault type occurring. Step 5: After processing the real-time collected basic data in Steps 2 and 3, the sensor anomaly score and entity vector are obtained and input into the graph neural network to generate the predicted probability of each fault type. When the predicted probability of any fault type is greater than a preset threshold, a fault warning signal is output. After forming a structured knowledge graph, Granger causality analysis is used to determine the strength of the causal association between sensor feature data and fault events, and the knowledge graph is updated, including the following steps. Historical sensor measurement data and fault events at the same time are collected to form a time series data set, where a fault event refers to the occurrence of a certain fault. For each type of sensor data and the fault events in historical records, a Granger causality model is established, represented as follows: ； In the formula, It refers to the probability that failure K will occur at time point t. Represents the intercept term. It is a constant. The fault autoregressive coefficient, Indicates the lag time of the i-th sensor channel. The influence coefficient on the fault, where L is the lag order. This refers to the time t- of the i-th sensor. The observed values, This refers to the fault type k at time t- The state; The optimal lag order L is dynamically determined using the Akaike Information Criterion, and the regression results are analyzed accordingly. The coefficient sequences are subjected to joint tests, the statistic is calculated, and a significance threshold is set. If the significance is satisfied, it is determined that the feature data xi has a significant causal relationship with the fault event. The confirmed causal relationships are explicitly added to the knowledge graph.

2. The equipment fault prediction method based on knowledge graph as described in claim 1, characterized in that: The score is a linear combination of the sensor anomaly score and the weights updated after message propagation between nodes.

3. The equipment fault prediction method based on knowledge graph as described in claim 2, characterized in that: The basic data includes voltage, current, temperature, and vibration amplitude; the historical fault data of the equipment includes equipment fault repair records and alarm logs; and the equipment topology and attribute data includes the hierarchical structure, connection relationships, and specification parameters of the equipment and its components. In a knowledge graph, defining entity nodes and relation edges means defining entity nodes that represent devices, components, sensor measurement points, and fault types, as well as relation edges that represent subordinate relationships, monitoring relationships, and fault association relationships.

4. The equipment fault prediction method based on knowledge graph as described in claim 3, characterized in that: The cosine similarity represents the vector direction correlation between sensor features and fault types, and the Euclidean distance represents the spatial distance between sensor features and fault type vectors.

5. The equipment fault prediction method based on knowledge graph as described in claim 4, characterized in that: Preprocessing the historical fault data of the equipment includes using natural language processing methods to extract fault entities, fault types and occurrence timestamps from equipment fault maintenance records and alarm logs, and generating structured fault event data; Preprocessing the device topology and attribute data includes extracting and generating structured topology data that characterizes the hierarchical structure, connection relationships, and specification parameters of the device and its components; The preprocessing of the basic data includes bandpass filtering and normalization of voltage, current and vibration amplitude data, sliding mean filtering and normalization of temperature data, and extraction of trend features within the sliding window, including the sliding window mean, standard deviation and trend slope.

6. The equipment fault prediction method based on knowledge graph as described in claim 5, characterized in that: The joint mapping features constituting each fault type include, By combining outlier scores, Euclidean distance, and cosine similarity, a joint weighted contribution is constructed, denoted as: ； In the formula, For sensor anomaly score, , These are the weighting coefficients. For cosine similarity, The distance is Euclidean. For each fault type node By aggregating the contributions from all related sensors, their joint mapping features are obtained, denoted as: ； In the formula, m represents the number of sensors. This is the embedding vector of the sensor node.

7. The equipment fault prediction method based on knowledge graph as described in claim 6, characterized in that: The score is represented as a linear combination of the sensor anomaly score and the weights updated by inter-node message propagation. ； In the formula, As weight, For sensor anomaly scores, The sensor anomaly score at time t; The probability of each fault type occurring is expressed as follows: ； In the formula, To predict the type of failure that will occur at time t The probability is given by K, where K is the total number of fault types and j is the traversal parameter for the fault types.

8. The equipment fault prediction method based on knowledge graph as described in claim 1, characterized in that: After constructing the graph neural network, the weight parameters of the graph neural network are updated using the cross-entropy loss function, based on historical equipment operation data and fault event labels.

9. A knowledge graph-based equipment failure prediction system, based on the knowledge graph-based equipment failure prediction method according to any one of claims 1 to 8, characterized in that: include, The map construction module is used to collect historical fault data of equipment, as well as equipment topology and attribute data, and to perform structured processing on the collected data. The feature extraction module is used to collect and process basic data during equipment operation. The vector mapping module is used to generate node vector representations based on the structural relationships between nodes in the knowledge graph and to achieve feature fusion. The fault reasoning module is used to build a graph neural network model based on the knowledge graph and to perform fault type prediction and reasoning. The real-time early warning module is used to receive real-time collected basic data and perform fault prediction based on the trained graph neural network.

Images