Knowledge graph-based industrial large model fault diagnosis method and system

Abstract

The application relates to the field of fault diagnosis, and discloses an industrial large model fault diagnosis method and system based on a knowledge graph, which is used for realizing zero-sample generalization capability for new fault modes in engineering. The method comprises the following steps: extracting fault nodes and observation nodes from equipment design documents and fault mode influence analysis tables, constructing an industrial knowledge graph, giving a uniform unit weight to each edge, performing reverse breadth-first search on each observation node to obtain a response index table, collecting sensor data in real time to obtain an abnormality degree vector, supporting a timestamp alignment and a rollback mechanism when MAD is zero, obtaining an initial score vector, obtaining a confidence degree and sorting and outputting Top-N faults, and outputting a shortest causal chain node sequence corresponding to each fault as a diagnosis basis. Through directional causal modeling, robust anomaly detection and sparse calculation, the application realizes efficient and interpretable industrial fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis, and in particular to a method and system for fault diagnosis of large industrial models based on knowledge graphs. Background Technology

[0002] With the deepening of Industry 4.0 and intelligent manufacturing, the structure of large industrial equipment is becoming increasingly complex, and failure modes are characterized by multi-factor coupling and concealed propagation paths. As a core link in ensuring the safe operation of equipment, the accuracy and real-time performance of fault diagnosis directly affect production efficiency and safety levels.

[0003] In recent years, fault diagnosis methods based on knowledge graphs have gradually become a research hotspot. These methods provide an interpretable knowledge foundation for diagnostic reasoning by structuring and storing equipment fault modes, sensor observations, and physical causal relationships. Meanwhile, the application of large language models in industrial fields has also promoted a fusion diagnostic paradigm of knowledge graphs and large models, attempting to combine the structured reasoning capabilities of knowledge graphs with the semantic understanding capabilities of large models.

[0004] Existing knowledge graphs are mostly constructed using undirected graphs or co-occurrence relationships, lacking a clear causal propagation direction between faulty nodes and observed nodes. This makes it impossible to distinguish whether a fault actually causes an observed anomaly during diagnosis, forcing reliance on statistical correlation for inference, resulting in a high false alarm rate and difficulty in providing a clear causal explanation to operations and maintenance personnel.

[0005] In industrial settings, different sensors have different sampling frequencies and inconsistent timestamps. Existing methods often ignore time alignment and directly use raw data to calculate anomalies, which disrupts the temporal logic of causal propagation and leads to temporal inconsistencies in diagnostic results.

[0006] Existing methods output a fault label or a Top-N list, but do not provide the reasoning behind the fault. Maintenance personnel cannot verify the diagnostic logic, making it difficult to build trust and develop targeted maintenance strategies.

[0007] When the number of sensors is large, the anomaly degree of all observation nodes is included in the calculation. Tiny sensor drifts can be accumulated and amplified, introducing a large amount of noise. Existing methods lack a filtering mechanism for low-confidence anomalies.

[0008] Therefore, we propose a knowledge graph-based industrial large-scale model fault diagnosis method and system to solve the above problems. Summary of the Invention

[0009] This invention provides a knowledge graph-based industrial large-scale model fault diagnosis method and system, which enables zero-sample generalization of new fault modes in engineering.

[0010] The first aspect of this invention provides a fault diagnosis method for large industrial models based on knowledge graphs. The method includes: constructing a knowledge graph containing fault nodes, observation nodes, and directed edges connecting the fault nodes and the observation nodes; determining the path length from the fault node to the observation node based on the knowledge graph, calculating a response coefficient based on the path length, and constructing a response index table based on the response coefficient; acquiring real-time monitoring data and historical monitoring data of the observation nodes, calculating the anomaly degree of the real-time monitoring data based on the historical monitoring data, and constructing an anomaly degree vector based on the anomaly degrees of all the observation nodes; combining the anomaly degree vector with the response index table to calculate an initial score for each fault node, forming an initial score vector; determining the confidence level of each fault node based on the initial score vector, and outputting a diagnosis result based on the confidence level.

[0011] Optionally, in a first implementation of the first aspect of the present invention, the directed edges are directed causal edges, and each directed causal edge has a set weight to form a causal chain structure; the knowledge graph is constructed by extracting fault modes from the equipment design document and establishing causal relationships between the fault nodes and the observation nodes.

[0012] Optionally, in a second implementation of the first aspect of the present invention, the method includes: performing a reverse breadth-first search starting from the observation node, recording the shortest causal chain length leading to the faulty node as the path length; taking the reciprocal of the shortest causal chain length as the response coefficient of the faulty node to the observation node, and organizing it into a response index table.

[0013] Optionally, in a third implementation of the first aspect of the present invention, during the execution of the reverse breadth-first search: when the same fault node is reached via different branch paths and the path lengths are equal, the reciprocal of the path length is taken and divided by the total number of paths of equal length to obtain the response coefficient; when the same fault node is reached but the path lengths are different, only the shortest path length is retained to calculate the response coefficient.

[0014] Optionally, in the fourth implementation of the first aspect of the present invention, the method includes: acquiring the historical monitoring data through a sliding window, calculating the median and median absolute deviation of the historical monitoring data; dividing the absolute value of the deviation of the real-time monitoring data from the median by the median absolute deviation to obtain a calculation result, and truncating the calculation result to a preset range as the anomaly degree.

[0015] Optionally, in a fifth implementation of the first aspect of the present invention, when calculating the anomaly degree: when the absolute deviation of the median is zero, the absolute value is divided by twice the standard deviation of the historical monitoring data and truncated to the preset range as the anomaly degree; when the standard deviation is zero, the anomaly degree is directly assigned to zero.

[0016] Optionally, in a sixth implementation of the first aspect of the present invention, the method includes: replacing elements in the anomaly vector that are less than a preset threshold with zero to obtain a sparse anomaly vector; using the sparse anomaly vector as a row vector, performing a matrix-vector dot product operation with the response index table, and summing the corresponding product results to obtain the initial score of the fault node.

[0017] Optionally, in a seventh implementation of the first aspect of the present invention, virtual fault energy is defined as the initial score, and its calculation formula is as follows: ; In the formula, Let be the initial score of the i-th faulty node; M is the total number of observed nodes; For response coefficients; This represents the value of the j-th element in the sparse anomaly vector; It is the equivalent of the basic failure index.

[0018] Optionally, in the eighth implementation of the first aspect of the present invention, the monitoring data of the observation node is acquired by a sensor, and the method further includes performing timestamp alignment processing on the sensor data: acquiring the original sensor data stream and the corresponding original sampling timestamp; selecting a unified time axis as a reference, and using the original sampling points corresponding to the original sampling timestamps to perform interpolation calculations to obtain interpolated data at the unified timestamp position, forming an aligned sensor dataset, wherein the aligned sensor dataset is used to provide the historical monitoring data and the real-time monitoring data.

[0019] A second aspect of this invention provides a knowledge graph-based industrial large-scale model fault diagnosis system, comprising: a graph construction module for constructing a knowledge graph, the knowledge graph including fault nodes, observation nodes, and directed edges connecting the fault nodes and the observation nodes; an index construction module for determining the path length from the fault node to the observation node based on the knowledge graph, calculating a response coefficient based on the path length, and constructing a response index table based on the response coefficient; an anomaly vector module for acquiring real-time monitoring data and historical monitoring data of the observation nodes, calculating the anomaly degree of the real-time monitoring data based on the historical monitoring data, and constructing an anomaly degree vector based on the anomaly degrees of all the observation nodes; a score calculation module for calculating an initial score for each fault node by combining the anomaly degree vector and the response index table, forming an initial score vector; and a diagnosis output module for determining the confidence level of each fault node based on the initial score vector, and outputting a diagnosis result based on the confidence level.

[0020] The mechanism of this invention is as follows: prior causal knowledge in the industrial field is solidified into a structured response index table in the form of a unit weighted directed graph, so that the inverse of the length of the fault propagation path can be directly converted into a quantifiable response coefficient, thereby eliminating the dependence of traditional black box models on massive labeled data. Beneficial effects: By using the reciprocal of the shortest causal chain length as the response coefficient, the impact of faults on observations naturally decreases as the causal chain length increases. Compared to existing technologies that rely on expert experience to set weights or use equal weighting, this design unifies graph theory shortest path with diagnostic weights, avoiding the uncertainty of subjective parameter tuning and ensuring a reasonable prioritization of proximate faults over distant faults, significantly improving the physical interpretability of diagnostic results. By replacing the traditional mean-standard deviation scheme with the median absolute deviation, it is statistically immune to impulse noise and outliers, making it particularly suitable for non-Gaussian noise environments in industrial settings. Furthermore, a two-level fault-tolerance mechanism is designed: when the MAD is zero, it falls back to the standard deviation; when the standard deviation is also zero, it directly assigns a value, ensuring that the system will not crash due to division by zero under extreme data scenarios. By introducing an anomaly threshold filtering before matrix multiplication, low-confidence anomalies are set to zero before participating in the calculation, decoupling anomaly detection from inference computation. This prevents the full accumulation and amplification of minute sensor drift, effectively reducing the false alarm rate. Compared to existing full-scale calculation schemes, it maintains stable diagnostic accuracy even as the sensor scale increases. The response index table simultaneously produces response coefficients and the shortest causal chain node sequence in a single reverse BFS, enabling Top-N fault output to be accompanied by a complete reasoning path. Operation and maintenance personnel can directly verify the causal logic of faults, intermediate nodes, and observations, solving the core pain point of existing methods that only output fault labels and lack traceable evidence, and greatly improving human-machine trust and maintenance decision-making efficiency. When the same fault is reachable via multiple paths of equal length, the response coefficient is divided by the number of paths for equal distribution, rather than arbitrarily selecting a single path. This eliminates the arbitrariness of path selection and ensures that the response coefficient truly reflects the reachability of the fault rather than the search traversal order, thereby improving the objectivity and consistency of the quantification results. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of an embodiment of the industrial large-scale model fault diagnosis method based on knowledge graphs in this invention. Figure 2 This is a schematic diagram of another embodiment of the industrial large-scale model fault diagnosis method based on knowledge graphs in this invention; Figure 3 This is a schematic diagram of an embodiment of the knowledge graph-based industrial large model fault diagnosis system in this invention. Figure 4 This is a schematic diagram of an embodiment of the industrial large-scale model fault diagnosis device based on knowledge graphs in this invention. Detailed Implementation

[0022] This invention provides a knowledge graph-based industrial large-scale model fault diagnosis method and system, which achieves zero-sample generalization capability for new fault modes in engineering. The terms first, second, third, fourth, etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms include or have, and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] It needs to be clearly defined that the industrial big model mentioned in this invention refers to a large-scale, comprehensive industrial system causal network and mechanism architecture model that covers complex physical node relationships, aiming to achieve system-level reasoning through knowledge graphs combined with full sensor data, rather than specifically referring to the deep learning big language model in the field of artificial intelligence characterized by massive parameters.

[0024] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the industrial large-scale model fault diagnosis method based on knowledge graphs in this invention includes: 101. Construct an industrial knowledge graph. The industrial knowledge graph includes fault nodes, observation nodes, and directed causal edges connecting fault nodes and observation nodes. Each causal edge can be assigned a corresponding correlation coefficient based on historical fault probability or expert experience score (to facilitate the explanation of the graph algorithm mechanism in this embodiment, the most basic unified weight of 1 is used as the set weight example. In practice, non-unified weights can be used to distinguish strong and weak correlations), forming a causal chain structure.

[0025] It is understood that the executing entity of this invention can be a knowledge graph-based industrial large-scale model fault diagnosis system, or it can be a terminal or a server; the specific implementation is not limited here. This embodiment of the invention will be described using a server as an example.

[0026] It should be noted that the process is divided into three stages: defining entity nodes, tracing causal transmission, and assigning graph weights. From the mechanistic knowledge of industrial equipment, fault events are extracted as fault nodes, and sensor features are extracted as observation nodes. These are connected by directed edges based on the causal relationships within the physical system. Each directed edge is assigned a uniform weight of 1, resulting in a causal chain network.

[0027] The following is a specific implementation example of a centrifugal cooling water pump system in a chemical plant: Based on the mechanism of this water pump system, the physical nodes of the graph are constructed as follows: Fault nodes (3 in total): power supply module failure, lubricating oil pump failure, bearing burnout. Observation nodes (4 in total): bus voltage zeroing, low lubricating oil pressure, high bearing temperature, abnormal rotor vibration.

[0028] Based on the equipment failure propagation mechanism, directed causal edges are established, with a uniform weight set to 1: A power supply module failure (fault node) will cause the bus voltage to drop to zero (observation node), and a directed edge with a weight of 1 will be established.

[0029] A failure of the power supply module (fault node) will cause the lubricating oil pump to fail (fault node). A directed edge with a weight of 1 is established.

[0030] A failure of the lubricating oil pump (fault node) will cause low lubricating oil pressure in the pipeline (observation node). A directed edge is established with a weight of 1.

[0031] A failure of the lubricating oil pump (fault node) will cause the equipment to lose lubrication and cooling, leading to the burning of the bearing (fault node). A directed edge is created with a weight of 1.

[0032] If a bearing burns out (fault node), it will be monitored as a high bearing temperature (observation node), and a directed edge will be created with a weight of 1.

[0033] A burnt-out bearing (fault node) will cause abnormal rotor vibration (observation node). A directed edge with a weight of 1 is established.

[0034] The aforementioned nodes and edges constitute a directed industrial knowledge graph. The graph forms causal chains of varying lengths, leading from power module failure to lubricating oil pump failure, then to bearing burnout, and finally to the high bearing temperature. Each edge has a uniform weight, providing a basis for calculating the shortest causal chain length.

[0035] 102. Based on the industrial knowledge graph, perform a reverse breadth-first search for each observation node, with the search depth limited to the preset maximum causal chain length. Record the shortest causal chain length from each fault node to each observation node. Take the reciprocal of each shortest causal chain length as the response coefficient of the fault node to the observation node. Summarize the response coefficients between all fault nodes and observation nodes into a response index table. The response index table has fault nodes as rows, observation nodes as columns, and table entries as response coefficients.

[0036] It should be noted that a graph algorithm is used to quantify the impact of faulty nodes on observed nodes and obtain an index table. The operations are as follows: setting the search depth, performing a reverse breadth-first search to obtain the shortest causal chain length, calculating the response coefficients, and summarizing to obtain the response index table.

[0037] Using the same cooling water pump system implementation example: The maximum causal chain length is preset to 3. During the search, only the fault propagation effect within 3 steps of the observation node is considered; the response coefficient outside this range is counted as 0.

[0038] For the four observation nodes, search in the opposite direction of the directed edge and record the shortest number of steps to reach the faulty node: 1. Reverse search from bus voltage zero: trace back to power supply module failure, with a minimum length of 1.

[0039] 2. Reverse search from low lubricating oil pressure: trace back to lubricating oil pump failure, the shortest length is 1; continue tracing to power supply module failure, the shortest length is 2.

[0040] 3. Starting from high bearing temperature, trace back to bearing burnout, with a minimum length of 1; continue tracing to lubricating oil pump failure, with a minimum length of 2; further trace to power supply module failure, with a minimum length of 3.

[0041] 4. Reverse search from rotor vibration anomaly: The topological location is parallel to the former, and the shortest lengths to trace back to bearing burnout, lubricating oil pump failure, and power supply module failure are 1, 2, and 3, respectively.

[0042] The reciprocal of the shortest causal chain length is used to convert it into response coefficients (length 1 is 1.00, length 2 is 0.50, length 3 is 0.33, and unreachable is 0.00). The summaries are shown in Response Index Table 1. Table 1

[0043] 103. Collect sensor data corresponding to each observation node in real time, collect historical data using a sliding window, calculate the median of sensor data within the sliding window and the absolute deviation of the median of sensor data within the sliding window, divide the absolute value of the current sensor observation value from the median by the absolute deviation of the median to obtain the current anomaly degree and truncate it to between 0 and 1. The current anomaly degrees of all observation nodes constitute an anomaly degree vector.

[0044] It should be noted that a sliding window is used to calculate the median and median absolute deviation of the sensor data, and a zero-prevention logic is introduced. Then, the ratio of the absolute value of the current value's deviation from the median to the absolute deviation is calculated, and the resulting vector is truncated to obtain the anomaly vector. The sliding window size is set to 5. To prevent the divisor from being zero, a default rule is set: if the calculated absolute deviation of the median is zero, then the minimum resolution of the sensor is used as the divisor instead.

[0045] 1. Bus voltage: Historical data: 379V, 380V, 380V, 381V, 382V. The median is 380V. The absolute deviation from the median (i.e., the absolute deviation of the median) is 1V.

[0046] 2. Lubricating oil pressure: Historical data: 0.49 MPa, 0.50 MPa, 0.51 MPa, 0.50 MPa, 0.49 MPa. The median is 0.50 MPa. The absolute deviation of the median is 0.01 MPa.

[0047] 3. Bearing temperature: Historical data: 44℃, 45℃, 46℃, 45℃, 44℃. The median is 45℃. The absolute deviation of the median is 1℃.

[0048] 4. Rotor vibration: Historical data: 1.8 mm / s, 2.0 mm / s, 2.2 mm / s, 2.0 mm / s, 1.8 mm / s. Median: 2.0 mm / s. Median absolute deviation: 0.2 mm / s.

[0049] Suppose that the bearing experiences an abnormal failure at the current moment, causing a significant increase in temperature, accompanied by abnormal vibration: 1. Bus voltage: Current value 380V. The absolute deviation is 0, divided by the deviation of 1, the result is 0. After truncation, the current anomaly degree is 0.

[0050] 2. Lubricating oil pressure: Current value 0.50 MPa. The absolute deviation is 0, divided by the deviation of 0.01, the result is 0. The current anomaly degree after truncation is 0.

[0051] 3. Bearing temperature: The current value has surged to 85°C. The absolute value of the deviation from the median of 45 is 40. Dividing 40 by the deviation of 1 gives 40. Truncate to between 0 and 1, the current anomaly is 1.

[0052] 4. Rotor vibration: Current value is 2.1 mm / s. The absolute deviation is 0.1. Dividing 0.1 by the deviation of 0.2 yields a result of 0.5. The current anomaly level is 0.5.

[0053] The current anomaly scores of the four observation nodes are combined to obtain the current anomaly score vector: [0,0,1,0.5].

[0054] 104. Perform a matrix-vector dot product operation on the anomaly vector and the response index table. Specifically, for each fault node, multiply all the response coefficients of its corresponding row in the response index table with the current anomaly of each observed node and sum them to obtain the initial score of the fault node. The initial scores of all fault nodes constitute the initial score vector.

[0055] It should be noted that the real-time anomaly vector is multiplied by the response index table using a matrix-vector dot product. The coefficient of each fault node's row is then multiplied by the anomaly vector item by item and summed to obtain the initial score.

[0056] Anomaly vector: [0,0,1,0.5] (corresponding to bus voltage, lubricating oil pressure, bearing temperature, and rotor vibration). Response index table: a table containing data for 3 fault nodes obtained in step 102.

[0057] Calculate the initial score for power supply module failure: Multiply the corresponding coefficients by the anomaly vector item by item: 1.00×0=0, 0.50×0=0, 0.33×1=0.33, 0.33×0.5=0.165; Add the above 4 results to get the initial score of 0.495.

[0058] Calculate the initial score for lubricating oil pump failure: multiply the corresponding coefficients and the anomaly vector item by item: 0.00×0=0, 1.00×0=0, 0.50×1=0.50, 0.50×0.5=0.25; add the above 4 results to get an initial score of 0.75.

[0059] Calculate the initial score for bearing burnout: multiply the corresponding coefficients and the anomaly vector item by item: 0.00×0=0, 0.00×0=0, 1.00×1=1.00, 1.00×0.5=0.50; add the above 4 results to get the initial score of 1.50.

[0060] The calculation results of the three fault nodes are combined to obtain the initial score vector: power supply module failure 0.495, lubricating oil pump failure 0.75, bearing burnout 1.50.

[0061] 105. Divide each term in the initial score vector by the preset absolute threshold of the device's historical maximum theoretical score (or map it using a probability distribution model) to obtain the confidence level of each fault node with absolute evaluative significance. When the confidence level exceeds the preset safety baseline, sort the fault nodes from high to low confidence levels and select the top few fault nodes as the diagnostic results output. Simultaneously, extract the node sequence from the shortest causal chain corresponding to each selected fault node from the response index table and output it as the diagnostic basis. (Note: The basis for extraction is the shortest causal chain corresponding to the observation node that actually generates a non-zero anomaly contribution. For example, even if there is a chain of length 1, if its observation node currently has anomaly of 0, it will not be used as the basis for this diagnosis. Instead, the longer causal chain that actually detected the anomaly will be output as inference support.) It should be noted that the process is divided into score normalization, confidence ranking and selection, and diagnostic criterion extraction. The initial score vector is divided by the maximum value to obtain the confidence level. The top few nodes are then sorted by confidence level and output. The shortest causal chain sequence is then extracted from the graph as the diagnostic criterion.

[0062] The initial scores for the three nodes are known to be: power supply module failure 0.495, lubricating oil pump failure 0.75, and bearing burnout 1.50. The preset absolute score threshold for this system based on historical severe fault data is 2.00.

[0063] Calculate the confidence level by dividing each item by the absolute threshold of 2.00: Bearing burnout: 1.50 divided by 2.00, confidence level 0.75. Lubricating oil pump failure: 0.75 divided by 2.00, confidence level 0.375. Power supply module failure: 0.495 divided by 2.00, confidence level 0.2475.

[0064] Sort by confidence level from highest to lowest: bearing burnout (0.75), lubricating oil pump failure (0.375), power supply module failure (0.2475). The first two nodes are selected by default, i.e., bearing burnout and lubricating oil pump failure are taken as the output results.

[0065] Based on the reverse search record in step 102, extract the shortest causal chain node sequence reaching the abnormal observation node: the shortest sequence for bearing burnout: sequence 1 (bearing burnout, high bearing temperature) and sequence 2 (bearing burnout, abnormal rotor vibration). The shortest sequence for lubricating oil pump failure: sequence 1 (lubricating oil pump failure, bearing burnout, high bearing temperature) and sequence 2 (lubricating oil pump failure, bearing burnout, abnormal rotor vibration).

[0066] The summarized fault diagnosis report is shown in Table 2 below: Table 2

[0067] Please see Figure 2 Another embodiment of the knowledge graph-based industrial large-scale model fault diagnosis method in this invention includes: 201. Construct an industrial knowledge graph. The industrial knowledge graph includes fault nodes, observation nodes, and directed causal edges connecting fault nodes and observation nodes. Each causal edge can be assigned a corresponding correlation coefficient based on historical fault probability or expert experience score (to facilitate the explanation of the graph algorithm mechanism in this embodiment, the most basic unified weight of 1 is used as the set weight example. In practice, non-unified weights can be used to distinguish strong and weak correlations), forming a causal chain structure.

[0068] Specifically, constructing an industrial knowledge graph includes: extracting the fault nodes corresponding to the equipment and the observation nodes corresponding to the sensors from the equipment design documents and the failure mode effect analysis table; establishing directed causal edges between each fault node and each observation node based on the physical connection relationships recorded in the equipment design documents and the causal propagation directions recorded in the failure mode effect analysis table; assigning the same unit weight value to each directed causal edge; and storing all nodes and directed causal edges in the form of an adjacency list to obtain the industrial knowledge graph.

[0069] It should be noted that the following is a specific implementation example of a 2MW wind turbine gearbox system in a wind farm: Engineers reviewed the equipment design documents and historical FMEA records for this gearbox model, extracting specific fault points and observation points: Fault points (3 in total): Lubricating oil pump failure, high-speed shaft gear pitting, and high-speed bearing damage. Observation points (4 in total): Low oil pressure in the main lubricating oil pipeline (corresponding to the pressure transmitter), high gearbox return oil temperature (corresponding to the PT100 temperature sensor), large radial vibration amplitude (corresponding to the radial acceleration sensor), and large axial vibration amplitude (corresponding to the axial displacement sensor).

[0070] Based on physical connections and fault evolution directions, directed causal edges are established, with a uniform weight of 1: The lubricating oil pump stoppage directly leads to a loss of oil supply pressure in the pipeline. A directed edge is established from the lubricating oil pump stoppage to low oil pressure in the main lubricating oil pipeline, with a weight of 1. The oil pump stoppage causes the high-speed shaft system to lose its lubricating oil film, leading to abnormal wear. Two directed edges are established respectively from the lubricating oil pump stoppage to pitting corrosion of the high-speed shaft gears and damage to the high-speed bearings, each with a weight of 1. After pitting corrosion on the gear surface, friction generates heat, which is carried away by the return oil. A directed edge is established from pitting corrosion of the high-speed shaft gears to high return oil temperature in the gearbox, with a weight of 1. Pits in the gears disrupt the meshing state, causing radial impact. A directed edge is established from pitting corrosion of the high-speed shaft gears to large radial vibration amplitude, with a weight of 1. Bearing damage generates severe frictional heat. A directed edge is established from high-speed bearing damage to high return oil temperature in the gearbox, with a weight of 1. Bearing damage causes axial movement of the rotor. A directed edge is established from high-speed bearing damage to large axial vibration amplitude, with a weight of 1.

[0071] The above network is transformed into an adjacency list data structure, with the starting node as the index to record the directly pointed-to neighbor nodes and edge weights: [Lubricating oil pump stoppage] indicates: [Low oil pressure in main lubricating oil pipeline] (weight 1), [Pitting of high-speed shaft gear] (weight 1), [Damage to high-speed bearing] (weight 1).

[0072] [High-speed shaft gear pitting] indicates: [High gearbox return oil temperature] (weight 1), [Large radial vibration amplitude] (weight 1).

[0073] [High-speed bearing failure] indicates: [High gearbox oil return temperature] (weight 1), [Large axial vibration amplitude] (weight 1).

[0074] Each observation node serves as the end of a causal chain, and the entries it points to are all empty sets. This structure preserves the topological structure and provides a basis for reverse search.

[0075] 202. Based on the industrial knowledge graph, perform a reverse breadth-first search for each observation node, with the search depth limited to the preset maximum causal chain length. Record the shortest causal chain length from each fault node to each observation node. Take the reciprocal of each shortest causal chain length as the response coefficient of the fault node to the observation node. Summarize the response coefficients between all fault nodes and observation nodes into a response index table. The response index table has fault nodes as rows, observation nodes as columns, and table entries as response coefficients.

[0076] Specifically, performing a reverse breadth-first search includes: starting from each observation node, performing a breadth-first search in the reverse direction of the directed causal edge, and setting the upper limit of the search depth to the preset maximum causal chain length; during the search process, when a fault node is encountered for the first time, recording the shortest causal chain length from the observation node to the fault node and the sequence of nodes traversed by the shortest causal chain, and taking the reciprocal of the shortest causal chain length as the response coefficient of the fault node to the observation node; organizing the response coefficients between all fault nodes and observation nodes and the corresponding shortest causal chain node sequences into a two-dimensional table with fault nodes as rows and observation nodes as columns, each table entry containing the response coefficient of the fault node in that row to the observation node in that column and the corresponding shortest causal chain node sequence, thus obtaining a response index table.

[0077] Furthermore, during the reverse breadth-first search process, a set of visited nodes is maintained for each observation node. When the reverse search encounters a node that already exists in the set of visited nodes, the node and its subsequent branches are skipped, and the search is not repeated. When the same observation node starts from the same observation node, reaches the same fault node through different reverse branch paths, and all paths have the same length, the reciprocal of the length is taken and divided by the total number of paths of the same length to obtain the response coefficient of the fault node to the observation node. The node sequence is then extracted from any one of these paths as the shortest causal chain node sequence corresponding to the fault node. When the same fault node is reached but the path lengths are different, only the path corresponding to the shortest length and its node sequence are retained, and the response coefficient is the reciprocal of the shortest length.

[0078] It should be noted that by introducing an anti-dead-loop mechanism for the visited node set and a multi-path equal-length decay mechanism, a restricted reverse breadth-first search is performed to obtain a two-dimensional index table containing response coefficients and sequences. The wind turbine gearbox example is used as follows: The maximum causal chain length is preset to 3. Pathways exceeding 3 steps will be truncated, and the corresponding response coefficient will be counted as 0.

[0079] A search is performed on the four observed nodes based on the adjacency list, and the visited set is maintained: a reverse search is performed starting from low oil pressure in the main lubricating oil pipeline: tracing back to the point where the lubricating oil pump stops when the depth is 1. The shortest length is 1, recording the sequence [lubricating oil pump stops, low oil pressure in the main lubricating oil pipeline]. The reciprocal of the length is 1.00, and the response coefficient is 1.00.

[0080] Starting with a large radial vibration amplitude, the search proceeds backward: at a depth of 1, it traces back to pitting of the high-speed shaft gear, with a response coefficient of 1.00. Further, at a depth of 2, it traces back to the lubricating oil pump stopping, with a length of 2, a reciprocal of 0.50, and a response coefficient of 0.50. The corresponding sequence is recorded.

[0081] Starting with a large axial vibration amplitude, a reverse search is performed: a depth of 1 traces back to high-speed bearing failure, with a response coefficient of 1.00. A depth of 2 traces back to the lubricating oil pump stopping, with a response coefficient of 0.50.

[0082] The reverse search from high gearbox return oil temperature (including multi-path equal-length processing) shows that at depth 1, two reverse branches lead to pitting of the high-speed shaft gear and damage to the high-speed bearing, with a minimum length of 1 and a response coefficient of 1.00. At depth 2, both branches trace back to the lubricating oil pump stopping. The minimum length is now 2, with a reciprocal of 0.50. Two paths of equal length exist. Dividing the reciprocal 0.50 by the total number of paths (2) yields a response coefficient of 0.25. One sequence can be extracted: [lubricating oil pump stopping, high-speed shaft gear pitting, high gearbox return oil temperature].

[0083] The above search results are summarized into a two-dimensional table as shown in Table 3 below: Table 3

[0084] 203. Collect sensor data corresponding to each observation node in real time, collect historical data using a sliding window, calculate the median of sensor data within the sliding window and the absolute deviation of the median of sensor data within the sliding window, divide the absolute value of the current sensor observation value from the median by the absolute deviation of the median to obtain the current anomaly degree and truncate it to between 0 and 1. The current anomaly degrees of all observation nodes constitute an anomaly degree vector.

[0085] Specifically, the real-time acquisition of sensor data corresponding to each observation node includes: pre-setting a sliding window containing sensor data from several recent historical moments for each observation node, and acquiring the current sensor observation value in real time; extracting all historical data from the sliding window, and calculating the median and median absolute deviation of the historical data; dividing the absolute value of the difference between the current sensor observation value and the median by the median absolute deviation to obtain a ratio, and truncating this ratio to the range between 0 and 1, which is used as the anomaly degree of the observation node at the current moment; arranging the anomaly degrees of all observation nodes in the same order as the column order in the response index table to obtain the anomaly degree vector.

[0086] Furthermore, when calculating the anomaly of each observation node at the current moment, if all historical data in the sliding window are sorted and all values ​​are equal, resulting in a median absolute deviation of zero, the absolute value of the difference between the current sensor observation and the median is divided by twice the standard deviation of the data in the sliding window, and the resulting ratio is truncated to between 0 and 1, which is used as the anomaly of the observation node at the current moment; if the standard deviation of the data in the sliding window is also zero, the anomaly of the observation node at the current moment is directly assigned to 0.

[0087] It should be noted that the wind turbine gearbox embodiment is used. Based on the gradual change characteristics of fatigue wear, a sliding window is set to cache the daily on-time observation historical data for the most recent 5 days. Assuming that we are now in the 6th day, due to the deterioration of long-term latent gear pitting, a relatively significant increase in oil temperature and vibration has occurred: Low oil pressure point in the main lubricating oil pipeline: Historical data: The oil pressure for the past 5 days was 0.50MPa, 0.50MPa, 0.50MPa, 0.50MPa, 0.50MPa.

[0088] Indicator calculations: Median is 0.50 MPa, absolute deviation is 0, and standard deviation is also 0.

[0089] Current observation: The current value on day 6 is 0.50 MPa.

[0090] Anomaly calculation: Trigger the second fallback logic (standard deviation is 0), and the anomaly is directly assigned a value of 0.

[0091] High oil return temperature points in the gearbox: Historical data: Oil temperatures over the past 5 days were 59℃, 60℃, 60℃, 61℃, and 62℃.

[0092] Index calculation: The median is 60℃. The absolute values ​​of deviation from the median are 1℃, 0℃, 0℃, 1℃, and 2℃, respectively. The median of the absolute values ​​(i.e., the absolute deviation from the median) is 1℃.

[0093] Current observation: Day 6. The current temperature has soared to 65°C due to increased wear and tear.

[0094] Anomaly calculation: The absolute deviation is 5℃, divided by the deviation of 1℃ to get 5. The anomaly is 1.00 when the value is truncated to between 0 and 1.

[0095] Large radial vibration amplitude nodes: Historical data: Vibration velocities in the past 5 days were 1.8 mm / s, 1.9 mm / s, 2.0 mm / s, 2.1 mm / s, and 2.2 mm / s.

[0096] Index calculation: The median is 2.0 mm / s. The absolute deviations are 0.2 mm / s, 0.1 mm / s, 0 mm / s, 0.1 mm / s, and 0.2 mm / s. The absolute deviation of the median is 0.1 mm / s.

[0097] Current observation: Day 6, current value is 2.08 mm / s.

[0098] Anomaly calculation: The absolute deviation is 0.08 mm / s. Dividing by the deviation of 0.1 mm / s gives 0.8. No truncation is needed; the anomaly is 0.80.

[0099] Largest axial vibration amplitude nodes: Historical data: Data for the past 5 days are 1.4mm / s, 1.5mm / s, 1.5mm / s, 1.6mm / s, and 1.7mm / s.

[0100] Index calculation: Median is 1.5 mm / s. Absolute deviation of median is 0.1 mm / s.

[0101] Current observation: Day 6, current value is 1.5 mm / s (axial direction not affected by pitting corrosion).

[0102] Anomaly calculation: The absolute value of the deviation is 0. Dividing by the deviation of 0.1 mm / s gives 0, and the anomaly is 0.00.

[0103] The anomaly scores of the four nodes are combined in the order of the index columns to obtain the vector: [0, 1.00, 0.80, 0].

[0104] 204. Perform a matrix-vector dot product operation on the anomaly vector and the response index table. Specifically, for each fault node, multiply all the response coefficients of its corresponding row in the response index table with the current anomaly of each observed node and sum them to obtain the initial score of the fault node. The initial scores of all fault nodes constitute the initial score vector.

[0105] Specifically, performing a matrix-vector multiplication operation between the anomaly vector and the response index table includes: multiplying the anomaly vector as a row vector with the response index table, specifically: for each fault node in the response index table, multiplying the response coefficients corresponding to all entries in that row by the anomaly values ​​in the corresponding column positions of the anomaly vector, and then summing all the product results to obtain the initial score of the fault node; performing the above operation sequentially on all fault node rows, and arranging all the obtained initial scores according to the row order of the response index table to form an initial score vector.

[0106] Furthermore, before multiplying the anomaly vector with the response index table, each anomaly value in the anomaly vector is compared with a preset anomaly threshold. Anomaly values ​​below the threshold are replaced with 0 to obtain a sparse anomaly vector. Then, the sparse anomaly vector is used as a row vector to multiply with the response index table. Specifically, for each row of fault nodes in the response index table, only the entries in that row corresponding to the column positions of the non-zero items in the sparse anomaly vector are selected. The response coefficients corresponding to these entries are multiplied by the corresponding non-zero anomaly values ​​in the sparse anomaly vector. Then, all the product results are added together to obtain the initial score of the fault node. The above operation is performed sequentially for all fault node rows, and all the obtained initial scores are arranged in the row order of the response index table to form an initial score vector.

[0107] It should be noted that threshold filtering is used to eliminate interference, and matrix-vector mapping is completed using a calculation formula that is compatible with physical dimensions. The current anomaly vector is [0, 1.00, 0.80, 0].

[0108] Set the anomaly threshold to 0.50. Low lubricating oil main pipeline pressure: 0 (below 0.50), replace with 0. High gearbox return oil temperature: 1.00 (above 0.50), retain 1.00. Large radial vibration amplitude: 0.80 (above 0.50), retain 0.80. Large axial vibration amplitude: 0 (below 0.50), replace with 0. The sparse anomaly vector is determined as [0, 1.00, 0.80, 0].

[0109] Define a dimensionless comprehensive failure impact index as the initial score, and calculate it using the following formula:

[0110] in the formula Let be the initial score of the i-th faulty node; M is the total number of observed nodes (M=4). For response coefficients; It has a non-zero outlier. The basic failure index equivalent (here set as a dimensionless constant 1).

[0111] Extract the non-zero items from columns 2 and 3 for calculation: Lubricating oil pump stoppage score calculation: The response coefficient of this node at high gearbox return oil temperature (0.25) multiplied by the node's anomaly degree (1.00), plus the response coefficient of this node at large radial vibration amplitude (0.50) multiplied by the node's anomaly degree (0.80), results in 0.65. Substituting into the formula, we get S1 = 0.65.

[0112] Calculation of pitting score for high-speed shaft gears: Multiply the response coefficient of the node (1.00) under high gearbox return oil temperature by the node's anomaly degree (1.00), and add the response coefficient of the node under large radial vibration amplitude (1.00) multiplied by the node's anomaly degree (0.80), resulting in 1.80. Substituting into the formula, we get S2 = 1.80.

[0113] High-speed bearing damage score calculation: The response coefficient of this node at high gearbox return oil temperature (1.00) is multiplied by the node's anomaly degree (1.00), plus the response coefficient of this node at large radial vibration amplitude (0.00) multiplied by the node's anomaly degree (0.80), resulting in 1.00. Substituting into the formula, we get S3 = 1.00.

[0114] Arranged in row order, the initial score vector is [0.65, 1.80, 1.00].

[0115] 205. Divide each item in the initial score vector by the preset absolute threshold of the device's historical maximum theoretical score (or map it through a probability distribution model) to obtain the confidence level of each fault node with absolute evaluation significance. When the confidence level exceeds the preset safety baseline, sort the fault nodes from high to low confidence level and select the top few fault nodes as the diagnostic results output. At the same time, extract the node sequence in the shortest causal chain corresponding to each selected fault node from the response index table and output it as the diagnostic basis.

[0116] Specifically, normalizing the initial score vector by its maximum value includes: extracting the maximum value from all values ​​in the initial score vector; dividing the initial score of each fault node by this maximum value to obtain the confidence value of the fault node; forming a confidence vector by arranging all confidence values ​​according to the order of fault nodes in the initial score vector; sorting each value in the confidence vector from largest to smallest, while retaining the fault node identifier corresponding to each value, to obtain a sorted sequence of fault nodes; selecting a predetermined number of fault nodes from the sorted sequence of fault nodes according to the sorting order as selected fault nodes; for each selected fault node, reading the shortest causal chain node sequence contained in all entries of the corresponding row of the selected fault node from the response index table, and merging the confidence value of the selected fault node with all read shortest causal chain node sequences to form a diagnostic report entry for the selected fault node; summarizing the diagnostic report entries of all selected fault nodes according to the sorting order to obtain the diagnostic result and output it.

[0117] It should be noted that the absolute score at the energy level is mapped to a confidence level, and the structured diagnostic report is obtained by matching the spectral path.

[0118] The known scores are: lubricating oil pump failure 0.65, high-speed shaft gear pitting 1.80, and high-speed bearing damage 1.00. The default absolute score threshold for this system based on historical severe fault data is 2.00. High-speed shaft gear pitting: 1.80 divided by the absolute threshold of 2.00, confidence level 0.90 (90%); lubricating oil pump failure: 0.65 divided by 2.00, confidence level 0.325 (32.5%).

[0119] High-speed bearing failure: 1.00 divided by 2.00 absolute threshold, confidence level 0.50 (50%).

[0120] The confidence vector is [0.325, 0.90, 0.50].

[0121] The results, sorted from largest to smallest, are: high-speed shaft gear pitting (0.90), high-speed bearing damage (0.50), and lubricating oil pump failure (0.325). The first two nodes are selected as the default output.

[0122] Read the shortest causal chain node sequence corresponding to the selected nodes and summarize it to form Report Table 4: Table 4

[0123] 206. After constructing the industrial knowledge graph and before real-time acquisition of sensor data corresponding to each observation node, timestamp alignment is performed on the sensor data streams corresponding to each observation node: obtain all sensor data streams associated with each observation node and their original sampling timestamps, select a unified system time axis as the benchmark, and for each sensor data stream, use the linear interpolation method between adjacent original sampling points to calculate the interpolated data at each unified timestamp position, and organize the interpolated data of all observation nodes into an aligned sensor dataset according to the unified timestamp; the aligned sensor dataset is used for the construction of the sliding window and the acquisition of the current sensor observation values ​​in the real-time acquisition step.

[0124] Furthermore, during the timestamp alignment process, for each unified timestamp of each sensor data stream, the original sampling point closest to the unified timestamp is extracted from the original sampling points of the data stream, and the data value of the original sampling point is directly assigned to the interpolated data at the unified timestamp position to form the aligned sensor dataset. The data value at each unified timestamp position in the aligned sensor dataset is the data value in the original sampling point, which is used for the construction of the sliding window and the acquisition of the current sensor observation value in the real-time acquisition step.

[0125] It should be noted that, based on a unified long-period reference time axis, mathematical linear interpolation is used to eliminate the problems of different sampling frequencies and asynchronous arrival of different sensors, thereby reconstructing a structured equidistant historical dataset.

[0126] To match the gradual degradation characteristics of the equipment, the time axis is aligned to the reference by day, and the time of 12:00:00 of each day of the past 5 consecutive days is selected as the reference timestamp (i.e. noon of each day from day 1 to day 5).

[0127] Taking the alignment data of day 1 obtained in step 203 as an example, the original sampling data of each sensor hardware is distributed as follows. The system performs linear interpolation using the two closest original data points: For nodes with large radial vibration amplitude (Day 1, 12:00:00): First raw sampling point (before): Day 1, 08:00:00, value 1.70 mm / s.

[0128] Second raw sampling point (later): Day 1, 16:00:00, value 1.90 mm / s.

[0129] Linear interpolation calculation: Midpoint of the time interval, interpolation result is mm / s. This value is precisely aligned to 12:00:00.

[0130] For the high gearbox oil return temperature point (Day 1, 12:00:00): First point: Day 1, 06:00:00, 58.5℃.

[0131] Second point: Day 1, 18:00:00, 59.5℃.

[0132] The alignment result calculated by linear interpolation is 59.0℃.

[0133] For nodes with large axial vibration amplitude (Day 1, 12:00:00): First point: Day 1, 09:00:00, 1.35mm / s.

[0134] Second point: Day 1, 15:00:00, 1.45mm / s.

[0135] The alignment result calculated by linear interpolation is 1.40 mm / s.

[0136] For the low oil pressure point in the main lubricating oil pipeline (Day 1, 12:00:00): First point: Day 1, 11:00:00, 0.50 MPa.

[0137] Second point: Day 1, 13:00:00, 0.50MPa.

[0138] Linear interpolation remained stable, yielding a result of 0.50 MPa.

[0139] The system uses the same algorithm to continuously traverse and calculate the reference time of 12:00:00 for each day from day 2 to day 5.

[0140] Through linear mapping, the original asynchronous raw data stream with randomly scattered timestamps is reshaped into a matrix-ordered standard dataset, as shown in Table 5 below: Table 5

[0141] The above describes the knowledge graph-based industrial large-scale model fault diagnosis method in the embodiments of the present invention. The following describes the knowledge graph-based industrial large-scale model fault diagnosis system in the embodiments of the present invention. Please refer to [link / reference]. Figure 3An embodiment of the industrial large-scale model fault diagnosis system based on knowledge graph in this invention includes: a graph construction module 301, used to construct a knowledge graph, the knowledge graph including fault nodes, observation nodes, and directed edges connecting the fault nodes and the observation nodes; an index construction module 302, used to determine the path length from the fault node to the observation node based on the knowledge graph, calculate the response coefficient based on the path length, and construct a response index table based on the response coefficient; an anomaly vector module 303, used to acquire real-time monitoring data and historical monitoring data of the observation nodes, calculate the anomaly degree of the real-time monitoring data based on the historical monitoring data, and construct an anomaly degree vector based on the anomaly degree of all the observation nodes; a score calculation module 304, used to combine the anomaly degree vector and the response index table to calculate and obtain the initial score of each fault node, forming an initial score vector; and a diagnosis output module 305, used to determine the confidence degree of each fault node based on the initial score vector, and output the diagnosis result based on the confidence degree.

[0142] Figure 4 This is a schematic diagram of the structure of a knowledge graph-based industrial large-scale model fault diagnosis device according to an embodiment of the present invention. The device 400 may include: a processor 401, a receiver 402, a transmitter 403, and a memory 404. The receiver 402, transmitter 403, and memory 404 are respectively connected to the processor 401 via a bus. It should be noted that in some possible implementations, the processor 401 and the memory 404 may be integrated together.

[0143] The processor 401 includes one or more processing cores. The processor 401 executes the methods performed by the base station in the random access method provided in this application embodiment by running software programs and modules. The memory 404 can be used to store software programs and modules. Specifically, the memory 404 can store an operating system 4041 and at least one application module 4042 required for a function. The receiver 402 is used to receive communication data sent by other devices, and the transmitter 403 is used to send communication data to other devices.

[0144] The present invention also provides a knowledge graph-based industrial large-scale model fault diagnosis device, which includes a memory and a processor. The memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor performs the steps of the knowledge graph-based industrial large-scale model fault diagnosis method in the above embodiments.

[0145] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the knowledge graph-based industrial large model fault diagnosis method.

[0146] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0147] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0148] The above-described 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A knowledge graph-based industrial large model fault diagnosis method, characterized in that, Includes the following steps: Construct a knowledge graph, which includes fault nodes, observation nodes, and directed edges connecting the fault nodes and the observation nodes; The path length from the faulty node to the observed node is determined based on the knowledge graph, the response coefficient is calculated based on the path length, and a response index table is constructed based on the response coefficient. Obtain real-time monitoring data and historical monitoring data of the observation nodes, calculate the anomaly degree of the real-time monitoring data based on the historical monitoring data, and construct an anomaly degree vector based on the anomaly degree of all the observation nodes. By combining the anomaly vector with the response index table, an initial score is obtained for each fault node, forming an initial score vector; The confidence level of each fault node is determined based on the initial score vector, and the diagnostic result is output according to the confidence level.

2. The industrial large-scale model fault diagnosis method based on knowledge graphs according to claim 1, characterized in that, The directed edges are directed causal edges, and each directed causal edge has a set weight to form a causal chain structure; The knowledge graph is constructed by extracting fault modes from equipment design documents and establishing causal relationships between the fault nodes and the observation nodes.

3. The industrial large-scale model fault diagnosis method based on knowledge graphs according to claim 2, characterized in that, include: Perform a reverse breadth-first search starting from the observation node, and record the shortest causal chain length to the faulty node as the path length; The reciprocal of the shortest causal chain length is used as the response coefficient of the faulty node to the observed node, and a response index table is formed.

4. The industrial large-scale model fault diagnosis method based on knowledge graphs according to claim 3, characterized in that, During the execution of the reverse breadth-first search: When the same fault node is reached via different branch paths and the path lengths are equal, the reciprocal of the path length is taken and divided by the total number of paths of equal length to obtain the response coefficient. When the same fault node is reached but the path lengths are different, only the shortest path length is retained to calculate the response coefficient.

5. The industrial large-scale model fault diagnosis method based on knowledge graphs according to claim 1, characterized in that, include: The historical monitoring data is obtained through a sliding window, and the median and median absolute deviation of the historical monitoring data are calculated. The absolute value of the deviation of the real-time monitoring data from the median is divided by the absolute deviation of the median to obtain the calculation result, and the calculation result is truncated to a preset range as the anomaly degree.

6. The industrial large-scale model fault diagnosis method based on knowledge graphs according to claim 5, characterized in that, When calculating the anomaly degree: When the absolute deviation of the median is zero, the absolute value is divided by twice the standard deviation of the historical monitoring data and truncated to the preset range, which is taken as the anomaly degree. When the standard deviation is zero, the anomaly degree is directly assigned to zero.

7. The industrial large-scale model fault diagnosis method based on knowledge graphs according to claim 1, characterized in that, include: Replace the elements in the anomaly vector that are less than a preset threshold with zero to obtain a sparse anomaly vector; The sparse anomaly vector is used as a row vector, and a matrix-vector dot product operation is performed with the response index table. The sum of the corresponding product results is used to obtain the initial score of the fault node.

8. The industrial large-scale model fault diagnosis method based on knowledge graphs according to claim 7, characterized in that, The virtual fault energy is defined as the initial score, and its calculation formula is as follows: ； In the formula, Let be the initial score of the i-th faulty node; M represents the total number of observation nodes; For response coefficients; This represents the value of the j-th element in the sparse anomaly vector; It is the equivalent of the basic failure index.

9. The industrial large-scale model fault diagnosis method based on knowledge graphs according to claim 1, characterized in that, The monitoring data from the observation nodes is acquired by sensors, and the method further includes timestamp alignment processing of the sensor data: Obtain the raw sensor data stream and the corresponding raw sampling timestamp; A unified time axis is selected as the reference, and interpolation calculations are performed using the original sampling points corresponding to the original sampling timestamps to obtain interpolated data at the unified timestamp positions, forming an aligned sensor dataset. The aligned sensor dataset is used to provide the historical monitoring data and the real-time monitoring data.

10. A fault diagnosis system for large industrial models based on knowledge graphs, characterized in that, include: The graph construction module is used to construct a knowledge graph, which includes fault nodes, observation nodes, and directed edges connecting the fault nodes and the observation nodes. An index building module is used to determine the path length from the fault node to the observation node based on the knowledge graph, calculate the response coefficient based on the path length, and build a response index table based on the response coefficient. An anomaly vector module is used to acquire real-time monitoring data and historical monitoring data of the observation nodes, calculate the anomaly degree of the real-time monitoring data based on the historical monitoring data, and construct an anomaly degree vector based on the anomaly degree of all the observation nodes. The scoring calculation module is used to combine the anomaly vector with the response index table to calculate the initial score of each fault node, forming an initial score vector. The diagnostic output module is used to determine the confidence level of each fault node based on the initial score vector, and output the diagnostic result according to the confidence level.

Images