A Knowledge Graph-Based Fault Location Method and System for Wind Turbine PLCs

Abstract

This invention provides a knowledge graph-based method and system for wind turbine PLC fault location. By analyzing the wind turbine physical design document and PLC control program source code, it effectively bridges the heterogeneous semantic gap. Based on this, it performs real-time dynamic event instantiation and graph embedding on the real-time operating data stream, generating dynamic node embedding vector sequences and dynamic activation subgraphs. This enables the knowledge graph to dynamically capture the wind turbine's operating status, overcoming the limitations of static topologies. An abnormal node is identified by introducing a logical-physical consistency propagation verification mechanism, and reverse causal path search and fault subgraph extraction are performed based on the dynamic activation subgraph. This accurately and interpretably locates the root cause of the fault, effectively mitigating information masking and false alarms caused by cascading effects. In this way, more accurate, faster, and smarter fault location of wind turbine PLCs can be achieved, significantly improving the operational reliability, maintenance efficiency, and overall economic benefits of wind turbine units.

Description

Technical Field

[0001] This invention relates to the field of intelligent positioning technology, specifically to a method and system for locating wind turbine PLC faults based on knowledge graphs. Background Technology

[0002] As a crucial component of clean energy, wind power relies heavily on the stable operation of its core equipment, the wind turbine, to ensure grid security and enhance power generation revenue. The programmable logic controller (PLC) is the core component connecting physical components and control strategies, responsible for real-time monitoring of the wind turbine's operating conditions, drive regulation, and safety protection. However, wind turbines operate for extended periods in complex outdoor environments, making them susceptible to extreme weather, component aging, sensor malfunctions, and other factors that can lead to PLC system failures. Therefore, developing an efficient and accurate method and system for locating wind turbine PLC faults is urgently needed to improve the overall operational reliability and economic efficiency of wind farms.

[0003] Traditionally, fault location methods for wind turbine PLCs have relied primarily on the experience of electrical engineers, manual troubleshooting, analysis of alarm history data, or single signal monitoring. These methods are often limited to specific domains, focusing either on physical-level fluctuation analysis or examining textual alarm logs. Existing technologies generally suffer from a core technical problem: a lack of physical-logic dual-layer coupling for fault tracing. Specifically, whether relying on physical signal vibrations or simply analyzing log text, traditional methods struggle to effectively align and verify the PLC's control logic flow with the physical equipment's operating status within a unified semantic space. This makes it difficult to diagnose complex and highly concealed faults such as "correct control logic but abnormal physical response" or "normal physical state but logical deadlock," severely limiting the depth and breadth of fault diagnosis.

[0004] Therefore, an optimized knowledge graph-based method for fault location in wind turbine PLCs is desired. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art, and provides a method and system for locating wind turbine PLC faults based on knowledge graphs.

[0006] In a first aspect, embodiments of the present invention provide a wind turbine PLC fault location method based on a knowledge graph, comprising: Obtain the wind turbine physical design document set and PLC control program source code; A two-layer static global knowledge graph is generated based on the wind turbine physical design document set and PLC control program source code. The acquired real-time wind turbine operation data stream is instantiated with real-time dynamic events and embedded into a graph to obtain a dynamic node embedding vector sequence and a dynamic activation subgraph. Logical-physical consistency propagation verification is performed on the dynamic node embedding vector sequence to obtain the set of abnormal nodes; Based on the dynamic activation subgraph, a reverse causal path search and fault subgraph extraction are performed on the abnormal node set to obtain the fault subgraph.

[0007] Secondly, embodiments of the present invention provide a wind turbine PLC fault location system based on a knowledge graph, comprising: The information acquisition module is used to acquire the wind turbine physical design document set and PLC control program source code; The knowledge graph generation module is used to generate a two-layer static global knowledge graph based on the wind turbine physical design document set and PLC control program source code. The real-time dynamic event instantiation and graph embedding module is used to perform real-time dynamic event instantiation and graph embedding on the acquired real-time wind turbine operation data stream to obtain dynamic node embedding vector sequences and dynamic activation subgraphs. The logical-physical consistency propagation verification module is used to perform logical-physical consistency propagation verification on the dynamic node embedded vector sequence to obtain the set of abnormal nodes. The fault subgraph extraction module is used to perform reverse causal path search and fault subgraph extraction on the set of abnormal nodes based on the dynamic activation subgraph to obtain the fault subgraph.

[0008] Compared with existing technologies, this invention provides a knowledge graph-based method and system for wind turbine PLC fault location. By parsing the wind turbine physical design document and PLC control program source code, it effectively bridges the heterogeneous semantic gap. Furthermore, it performs real-time dynamic event instantiation and graph embedding on the real-time operating data stream, generating dynamic node embedding vector sequences and dynamic activation subgraphs. This enables the knowledge graph to dynamically capture the wind turbine's operating status, overcoming the limitations of static topologies. Subsequently, by introducing a logical-physical consistency propagation verification mechanism to identify abnormal nodes, and performing reverse causal path search and fault subgraph extraction based on the dynamic activation subgraph, it accurately and interpretably locates the root cause of the fault, effectively mitigating the information masking and false alarm problems caused by cascading effects. This enables more accurate, faster, and smarter location of wind turbine PLC faults, significantly improving the operational reliability, maintenance efficiency, and overall economic benefits of wind turbine units. Attached Figure Description

[0009] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0010] Figure 1 This is a flowchart of a knowledge graph-based PLC fault location method for wind turbines according to an embodiment of the present invention; Figure 2 This is a data flow diagram of a knowledge graph-based PLC fault location method for wind turbines according to an embodiment of the present invention. Figure 3 This is a block diagram of a knowledge graph-based PLC fault location system for wind turbines according to an embodiment of the present invention. Detailed Implementation

[0011] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0012] Unless otherwise specifically stated, the technical or scientific terms used in the embodiments of this invention should be understood in their ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains. The terms "comprising" or "including," as used in the embodiments of this invention, do not limit the shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof mentioned, nor do they exclude the appearance or addition of one or more other different shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof, or the inclusion of these.

[0013] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale, and techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail; however, where appropriate, the illustrated techniques, methods, and apparatus should be considered part of the specification. In all the examples shown and discussed herein, any other specific example may have different values. It should be noted that similar symbols and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0014] In the description of the embodiments of the present invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In the embodiments of the present invention, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in the embodiments of the present invention, as well as the features of different embodiments or examples.

[0015] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0016] In the technical solution of this invention, a method for fault location of wind turbine PLC based on knowledge graph is proposed. Figure 1 This is a flowchart of a knowledge graph-based PLC fault location method for wind turbines according to an embodiment of the present invention. Figure 2 This is a system architecture diagram of a knowledge graph-based PLC fault location method for wind turbines according to an embodiment of the present invention. Figure 1 and Figure 2 As shown, the wind turbine PLC fault location method based on knowledge graph according to an embodiment of the present invention includes the following steps: S1, acquiring a wind turbine physical design document set and PLC control program source code; S2, generating a two-layer static global knowledge graph based on the wind turbine physical design document set and PLC control program source code; S3, performing real-time dynamic event instantiation and graph embedding on the acquired wind turbine real-time operation data stream to obtain a dynamic node embedding vector sequence and a dynamic activation subgraph; S4, performing logical-physical consistency propagation verification on the dynamic node embedding vector sequence to obtain an abnormal node set; S5, performing reverse causal path search and fault subgraph extraction on the abnormal node set based on the dynamic activation subgraph to obtain a fault subgraph.

[0017] Specifically, in step S1, the wind turbine physical design document set and the PLC control program source code are acquired. It should be understood that wind turbine faults often manifest as a mismatch between physical states and logical instructions. Relying solely on a single data source cannot construct a unified ontology encompassing cross-domain facts such as "variable address A controls solenoid valve B." In the technical solution of this invention, by simultaneously incorporating the static documents describing the physical topology and the source code describing the logical control flow into the system's view, fault tracing with a "physical-logic" dual-layer coupling can be achieved in subsequent steps. Specifically, the wind turbine physical design document set covers detailed information on all physical components of the wind turbine unit and their interrelationships, operating specifications, and performance parameters, such as mechanical structure diagrams, electrical schematics, sensor specifications, piping connection diagrams, hydraulic system principles, communication protocol specifications, and equipment operation manuals. These documents form the basis for understanding the physical structure of the wind turbine, the functions of its components, and their physical connections. They describe the static topology and physical properties of the equipment. The PLC control program source code is the core of defining the wind turbine's operating logic, typically written in programming languages ​​such as ladder diagrams, structured text, or function block diagrams. This source code details the mapping of input / output (I / O) points, the definition of internal variables, control algorithms, logical judgments, sequential control, safety interlock mechanisms, and alarm triggering conditions, which are crucial for understanding how the wind turbine self-regulates and protects itself under different operating conditions.

[0018] In practical implementation, this can be achieved through a specially configured information acquisition module. This module has multiple data interfaces to adapt to heterogeneous data from different sources. For acquiring the wind turbine physical design document set, the system establishes a connection with the wind farm's lifecycle management system (PLM) or document server to retrieve and download a complete set of engineering documents matching the model of the wind turbine being monitored. These documents not only contain mechanical drawings describing the equipment's appearance and dimensions, but more importantly, they cover piping and instrumentation diagrams (P&ID), electrical schematics, bills of materials (BOM), and sensor specifications describing component connections. For example, the system will focus on extracting drawings containing hydraulic station valve distribution, pitch motor wiring methods, and sensor installation locations. These documents define the physical boundaries of the wind turbine and the dependencies between entities. Simultaneously, for acquiring the PLC control program source code, this module connects to the wind turbine's main control system engineering station or version control server. The system will identify and extract control program files conforming to the IEC 61131-3 standard. These files are typically in the form of structured text (ST), ladder diagrams (LD), function block diagrams (FBD), or sequential function charts (SFC). The acquisition process is not limited to the main program file, but also includes related user-defined function blocks (UDFBs), global variable tables, and I / O mapping tables. It's worth noting that a version consistency check is performed during the acquisition process to ensure that the downloaded source code version matches the firmware version currently running on the wind turbine, preventing logical misjudgments due to version discrepancies.

[0019] Specifically, S2 generates a two-layer static global knowledge graph based on the wind turbine physical design document set and the PLC control program source code. It should be understood that existing technologies generally lack the ability to trace faults through physical-logical dual-layer coupling, making it difficult to effectively align and verify the PLC control logic flow and the operating status of physical equipment within a unified semantic space. In the technical solution of this invention, by constructing a two-layer static graph containing a physical topology layer and a control logic layer, the system can eliminate the heterogeneous semantic gap between the physical domain and the control domain, laying the foundation for subsequent realization of "physical-logical dual-layer coupling fault tracing capability," thereby enabling the identification and diagnosis of complex and highly concealed faults such as "correct control logic but abnormal physical response" or "normal physical state but logical deadlock."

[0020] In practical implementation, the first step is to construct a physical layer knowledge subgraph based on the wind turbine physical design document set. This involves parsing and extracting information from previously acquired physical design documents such as mechanical drawings, electrical schematics, P&ID diagrams, and component lists to identify the various physical entities of the wind turbine system (such as blades, gearboxes, generators, sensors, actuators, valves, and wiring) and their interrelationships in terms of structure, function, and connection. For example, the "inclusion" relationship between "hydraulic pump" and "hydraulic station," and the "connection" relationship between "oil pressure sensor" and "hydraulic pipeline" can be extracted from the piping and instrumentation diagram. These entities and relationships are abstracted as nodes and edges in the graph, collectively forming a knowledge subgraph describing the wind turbine's physical topology layer.

[0021] Next, based on the PLC control program source code, a logic layer knowledge subgraph is constructed. This means that the system performs lexical analysis, syntax analysis, and semantic analysis on the source code to identify function blocks, input / output (I / O) variables, internal memory addresses, data types, and the relationships between them in terms of calls, data flow, and control flow. For example, it can be identified that the function block "Pitch_Control_Module" calls the function block "Hydraulic_Valve_Drive," and that the variable "Pitch_Angle_Sensor_Input" is read by "Pitch_Control_Module." These logical entities and their programming relationships constitute the logic layer knowledge subgraph.

[0022] Furthermore, cross-domain semantic fusion is performed on the physical layer knowledge subgraph and the logical layer knowledge subgraph to obtain a two-layer static global knowledge graph. It should be understood that the wind turbine system itself is a typical mechatronic system, and its faults often manifest as physical-level anomalies interacting with the logical control layer through sensors or actuators. A purely physical layer knowledge graph cannot understand the PLC's control intent and logical flow, while a purely logical layer knowledge graph cannot perceive the actual state and structural limitations of the physical equipment. Therefore, in order to achieve comprehensive and accurate analysis of wind turbine faults, especially for logical-physical consistency propagation verification and reverse causal path search, the technical solution of this invention performs cross-domain semantic fusion on the physical layer knowledge subgraph and the logical layer knowledge subgraph to establish semantic associations between physical equipment and their corresponding PLC control variables and functional blocks, thereby providing a unified context for subsequent fault diagnosis. This cross-domain association is a fundamental prerequisite for achieving high-precision fault location.

[0023] In this process, firstly, entity anchor point alignment is performed on the physical layer knowledge subgraph and the logical layer knowledge subgraph to obtain a list of validated cross-domain entity pairs. This step aims to identify and match entities with the same semantics or direct relationships between the physical layer and logical layer knowledge subgraphs. For example, in the physical layer, a "gearbox oil temperature sensor" entity represents a specific physical component; while in the logical layer, there might be a PLC program block for "oil temperature acquisition" or a PLC variable named "GV_GearboxOilTemp," both closely related to the physical sensor. The entity anchor point alignment process identifies these corresponding entity pairs across physical and logical boundaries using automated or semi-automated methods (e.g., keyword matching based on entity names, semantic similarity calculation, or using predefined mapping rules). Once potential entity pairs are identified, they also need to be validated (e.g., through manual verification, domain expert rule validation, or consistency checks using graph structure features) to ensure that these alignments are accurate and meaningful, avoiding incorrect associations due to ambiguity or mismatches. Finally, the output is a list of cross-domain entity pairs that have passed the verification, which contains all physical-logical layer entity pairs that have been confirmed to have a corresponding relationship.

[0024] Next, cross-layer semantic edges are instantiated on the list of validated cross-domain entity pairs to obtain a set of cross-layer connection edges. That is, after obtaining the list of validated cross-domain entity pairs, the semantic relationships between these aligned entities are defined and represented as edges in a graph. For each cross-domain entity pair in the list, one or more semantic edges are instantiated based on its specific association type. For example, if a "physical motor entity" and a "PLC start command variable entity" are aligned, a semantic edge representing "controlled by" can be instantiated between them. These edges differ from connections within the physical layer or control flows within the logic layer; they are specifically designed to connect entities from different domains (physical and logical). The instantiation process of these semantic edges requires careful consideration of domain knowledge and engineering practices to ensure that the established connections accurately reflect the actual coupling relationship between the physical and logical aspects of the wind turbine system. All these created cross-layer semantic edges converge to form the final set of cross-layer connection edges.

[0025] Furthermore, a global heterogeneous graph data aggregation is performed on the physical layer knowledge subgraph, the logical layer knowledge subgraph, and the cross-layer connection edge set to obtain a two-layer static global knowledge graph. That is, all nodes (including physical entity nodes and logical variable / program nodes) and all edges (including physical connection edges, logical control flow edges, and newly created cross-layer semantic edges) are integrated into a single, unified graph structure. The resulting two-layer static global knowledge graph is a comprehensive static representation that fully reflects the physical topology of the wind turbine system and the PLC control logic and their interactions, providing a complete knowledge background for subsequent fault diagnosis and reasoning.

[0026] Specifically, in step S3, the acquired real-time wind turbine operation data stream is instantiated with real-time dynamic events and embedded into a graph to obtain a dynamic node embedding vector sequence and a dynamically activated subgraph. It should be understood that the real-time wind turbine operation data stream is typically a continuous, high-frequency sampled numerical sequence or state signal. This raw form of data cannot be directly reasoned about and analyzed on a static knowledge graph. Therefore, in the technical solution of this invention, the acquired real-time wind turbine operation data stream is instantiated with real-time dynamic events to transform this continuous data into discrete event tuples with clear semantics, such as "a parameter exceeds a threshold" or "equipment state switching." These events are dynamically changing, reflecting the behavior and state of the wind turbine at a specific moment. Secondly, in order to perform efficient computation and logical reasoning in the subsequent fault diagnosis stage, especially for complex graph structure analysis, these events and activated graph parts are further embedded into a graph to map the nodes and edges in the graph to a low-dimensional, continuous vector space. This embedding representation can capture the structural relationships and semantic information between nodes, enabling the computer to better understand and process complex knowledge graphs.

[0027] In practice, the system first performs edge detection and event tuple generation on the real-time operation data stream of the wind turbine to obtain an event tuple list. Edge detection refers to identifying key changes or anomalies in the data stream, such as sensor readings exceeding preset thresholds, PLC status bit flips, or significant anomalies in the data sequence detected through statistical methods. Once these edges are detected, the system converts them into structured event tuples. An event tuple typically contains information such as event type (e.g., high temperature warning, abnormal current), occurrence time, related device ID, and specific trigger value. For example, when the temperature data of a bearing on the wind turbine continuously rises and exceeds the safety threshold, the system generates an event tuple, such as (Event Type: 'Bearing Over-Temperature Alarm', Device ID: 'Main Bearing 1', Timestamp: '2023-10-26 10:30:15', Temperature Value: '85℃', Threshold: '80℃'). All these instantaneously generated event tuples are compiled into an event tuple list, which serves as the basis for subsequent dynamic activation.

[0028] Next, based on the list of event tuples, the two-layer static global knowledge graph is activated using event-driven dynamic logical subgraph activation to obtain a dynamically activated subgraph. That is, a local graph closely related to the current event is extracted from the pre-constructed two-layer static global knowledge graph; this is the dynamically activated subgraph. This process is event-driven, meaning that only knowledge graph parts directly or indirectly related to the real-time event will be activated.

[0029] In this process, firstly, the list of event tuples is traversed and indexed to obtain the set of activation seed nodes. Specifically, the system traverses each tuple in the event tuple list generated in the previous step. For each event tuple, it attempts to find a directly corresponding node in the two-layer static global knowledge graph. For example, for the "bearing over-temperature alarm" event, the system will look for a physical entity node representing "main bearing 1" or a node representing "bearing temperature sensor". These directly matched nodes constitute the set of activation seed nodes, which serve as the starting point for event activation.

[0030] Next, in the two-layer static global knowledge graph, a multi-hop logical propagation breadth-first search is performed on each activated seed node in the set of activated seed nodes to obtain the set of effective nodes and the set of effective edges. That is, starting from the set of activated seed nodes, a graph search algorithm is executed in the two-layer static global knowledge graph, specifically, along the edges of the graph to capture nodes logically or physically related to the seed nodes. For example, starting from the node "main bearing 1", the search can find its connected nodes such as "lubrication system", "temperature sensor", and possibly related nodes such as "PLC monitoring program block". To control the search range, a hop limit is usually set, i.e., only nodes within N hops of the seed node are searched. All nodes and edges visited during the search process together constitute the set of effective nodes and the set of effective edges.

[0031] Furthermore, the effective node set and effective edge set are dynamically focused on and reconstructed into a dynamic activation subgraph. That is, the effective node set and effective edge set are merged to construct the dynamic activation subgraph at the current moment. This subgraph represents the local knowledge graph that the system is most likely to focus on in the context of the current real-time event. It includes not only the devices and logic that caused the event, but also upstream and downstream components that are directly or indirectly related to them, thus providing a focused context for subsequent fault diagnosis.

[0032] Furthermore, spatiotemporal feature fusion and node embedding calculation are performed on the dynamic activation subgraph and the global node embedding vector matrix from the previous time step to obtain the dynamic node embedding vector sequence. In this process, firstly, the structural information of the current dynamic activation subgraph is fused with the temporal information in the global node embedding vector matrix from the previous time step. This can be achieved using techniques such as graph neural networks (GNNs), which can effectively aggregate node neighbor information and combine it with historical states. For example, recurrent neural networks (RNNs) or attention mechanisms can be used to handle temporal dependencies, ensuring that the embedding vector of the current node not only reflects its local structural context in the dynamic activation subgraph but also includes its historical evolution information.

[0033] Furthermore, based on the spatiotemporal feature fusion, the vector representation of each node in the dynamic activation subgraph is calculated, i.e., node embedding. These embedding vectors are low-dimensional, continuous numerical vectors that capture the node's position in the graph, its relationship with other nodes, and its dynamic changes over time. For example, for a node representing "motor current," its embedding vector not only contains information about the current value itself but may also incorporate contextual information such as its control variables, related sensors, and operating modes. Ultimately, all these node embedding vectors are arranged in sequence to form a dynamic node embedding vector sequence.

[0034] Specifically, in S4, a logical-physical consistency propagation verification is performed on the dynamic node embedding vector sequence to obtain the set of abnormal nodes. It should be understood that traditional wind turbine PLC fault location mechanisms use Euclidean distance to measure the difference between the logical prediction state and the physical feedback state. This model relies on a linear translation assumption, ignoring the complex nonlinear operating condition dependencies and anisotropic tolerance during wind turbine operation. Specifically, the response characteristics of wind turbine physical components are highly dependent on environmental conditions. For example, in low-temperature environments, increased hydraulic oil viscosity can cause valve lag. If only a fixed Euclidean distance threshold is used, such reasonable delays caused by changes in physical characteristics rather than logical faults will be incorrectly identified as signal deviations. Furthermore, different feature dimensions have drastically different sensitivities to errors. High-frequency oscillating signals such as instantaneous current allow for large variance fluctuations, while discrete state signals such as safety chain triggers require zero tolerance. The scalarized Euclidean distance forcibly treats all dimensions of error weights equally, causing the algorithm to be unable to distinguish between reasonable deviations under harsh operating conditions and genuine equipment fault deviations, thus leading to an extremely high false alarm rate under complex wind conditions. To address the aforementioned issues, this invention proposes an optimal mechanism that overcomes the limitations of traditional methods by introducing dynamic operating conditions and anisotropy tolerance. This mechanism enables precise and robust measurement of the deviation between logical expectations and actual physical feedback, thereby identifying truly inconsistent abnormal nodes with high confidence and providing an accurate starting point for subsequent root cause analysis.

[0035] In practice, the first step is to extract the first dynamic node embedding vector from the dynamic node embedding vector sequence. That is, in each verification cycle, the system sequentially selects the embedding vector corresponding to a node to be verified from the real-time generated dynamic node embedding vector sequence. This selected vector is called the first dynamic node embedding vector, representing the feature representation of the node's real-time running state at the current moment.

[0036] Next, the system extracts the environmental state node embedding set, logical-side expectation vector, and physical-side actual feedback node embedding vector associated with the first dynamic node embedding vector from the two-layer static global knowledge graph. Specifically, for the node corresponding to the currently selected first dynamic node embedding vector, the system uses the pre-constructed two-layer static global knowledge graph for querying and extraction: First, it extracts environmental factor nodes that the node depends on or is affected by, such as wind speed, ambient temperature, oil temperature, and power grid frequency, and obtains their corresponding embedding vectors to form the environmental state node embedding set. These embedding vectors provide current operating background information. Second, based on the node's logical context in the knowledge graph, such as its PLC function block or control program segment, it calculates or predicts the expected state or parameter value that the logic layer should present under the current operating conditions and represents it as the logical-side expectation vector. Then, it obtains the actual measurement / feedback values ​​of the physical sensors or actuators directly related to the node at the current moment and converts them into corresponding embedding vectors, forming the physical-side actual feedback node embedding vector, which reflects the actual situation in the physical world.

[0037] Subsequently, a dynamic operating condition context vector is synthesized from the first dynamic node embedding vector and the environmental state node embedding set to obtain the first operating condition context vector. This step aims to address the problem that a single control command cannot dynamically adapt to a changing environment. Specifically, by introducing a multi-head attention mechanism, the current control logic node embedding vector is used as the query term, and the associated environmental state node (such as wind speed, oil temperature, and grid frequency) embedding set is used as the key-value pair to calculate the nonlinear constraint weights of environmental factors on the control command. Based on this, the environmental context information is deeply integrated into the original control features using residual connections to generate a context-aware vector that can represent the control features that should be presented under the current specific operating condition; this process is expressed by the formula:

[0038] in, This represents the synthesized working condition context vector at time t; This is the embedding vector of the original control logic node; This is the feature matrix of the associated environmental nodes; and These are the projection weight matrices for the keys and values, respectively; This is the scaling factor; Presentation layer normalization operation.

[0039] Then, based on the first operating condition context vector, a first tolerance matrix is ​​generated. This step aims to construct an error tolerance ellipsoid that dynamically changes with the operating condition, thereby replacing the traditional fixed spherical threshold. Specifically, by inputting the operating condition context vector generated in the previous step into a multilayer perceptron network, a diagonal vector representing the allowable deviation range of each dimension is generated and converted into an inverse covariance matrix. The element values ​​in this matrix dynamically reflect the system's tolerance to deviations of different physical signals at the current moment. For example, when a high viscosity oil temperature context is detected, the weight value of the corresponding action delay dimension is automatically reduced, thereby giving the algorithm immunity to physical hysteresis and ensuring the flexibility and rationality of the fault judgment.

[0040] Furthermore, based on the first tolerance matrix, the expected vector on the logical side, and the actual feedback node embedding vector on the physical side, the violation score of the first dynamic node embedding vector is determined. That is, the consistency between logic and physics is ultimately determined based on the principle of statistical distance. Specifically, the quadratic distance between the difference vectors of the two is calculated under the weighted tolerance matrix. Deviations that conform to the physical characteristics of the current operating condition (i.e., differences falling within the low-weight region defined by the tolerance matrix) are automatically canceled out, while anomalies that violate physical laws or logical constraints (i.e., differences falling within the high-weight region) are significantly amplified. This achieves accurate online and offline differentiation between reasonable physical delays and true logical deadlocks, completing high-confidence root cause screening of faults.

[0041] Finally, based on the comparison between the violation score and a preset threshold, it is determined whether to include the node corresponding to the first dynamic node embedding vector in the abnormal node set. Specifically, if the violation score exceeds this threshold, it indicates a significant inconsistency between logic and physical behavior, and the node corresponding to the first dynamic node embedding vector and its associated behavior or state are identified as abnormal and added to the abnormal node set; conversely, if the violation score is within the threshold, it is considered to be within the normal or acceptable deviation range.

[0042] The aforementioned optimization mechanism, by constructing an anisotropic metric space capable of sensing physical context, achieves a leap from rigid numerical comparison to elastic semantic verification in wind turbine PLC fault diagnosis. Specifically, it eliminates the interference of inconsistent equipment responses caused by the complex environmental factors of wind farms on fault determination, effectively solving the problem that traditional methods cannot distinguish between normal physical hysteresis and equipment functional failure under high-dynamic wind conditions or extreme temperature conditions. Ultimately, this mechanism not only significantly reduces the false alarm rate caused by fluctuations in environmental conditions and improves the system's accuracy in identifying hidden logical faults, but also makes the fault location results physically interpretable, providing maintenance personnel with accurate decision support that integrates the operating condition context.

[0043] Specifically, in step S5, based on a dynamically activated subgraph, a reverse causal path search and fault subgraph extraction are performed on the set of abnormal nodes to obtain a fault subgraph. Since wind turbine systems are highly coupled, a fault point may propagate through physical connections or control logic, causing multiple downstream nodes to exhibit abnormalities. Focusing only on the abnormal node itself may lead to misjudgment. Therefore, in the technical solution of this invention, within the narrowed search space of the dynamically activated subgraph, a reverse causal path search is performed using the physical connections and logical control relationships defined in the graph. Through reverse tracing, the system can construct one or more causal chains pointing from symptoms to the root cause and extract them as a fault subgraph. This not only locates the source of the fault but also visualizes the fault propagation path, providing maintenance personnel with accurate decision support that is interpretable at both the physical and logical levels.

[0044] In practice, firstly, the dynamic activation subgraph output from the previous stage is received as the search scope, and the set of abnormal nodes is received as the starting point list for the search. Each node in the set of abnormal nodes is marked as a leaf node and placed into the processing queue. Next, for each anomalous node in the queue, the system performs a depth-first search (DFS) or breadth-first search (BFS) along the reverse direction of the edges in the dynamically activated subgraph. In a knowledge graph, edges typically represent data flow or control flow. In one example, if there is an edge A→B in the graph (indicating that A influences B or A controls B), and B is the current anomalous node, the search algorithm will trace back to A. Furthermore, the search process crosses the physical and logical boundaries. For example, if the current anomalous node is a PLC input variable in the logic layer, the algorithm will search backward along the cross-layer semantic edges to the sensor entity in the physical layer; if the current node is a controlled component in the physical layer, the algorithm will search backward to the PLC output instruction in the logic layer. During the traversal, the system records all visited nodes and edges, forming candidate causal paths.

[0045] Next, the searched predecessor nodes are relevance verified. This typically involves checking whether the predecessor node's state is also abnormal, or whether the predecessor node is in a specific state configuration that would cause downstream anomalies. In one example, if node B is abnormal and predecessor node A is found, the system will determine, according to predefined inference rules, whether A's state is sufficient to cause B's anomaly. For example, if A is a "stop command" and is in a TRUE state, then B (motor running state) being FALSE is reasonable; but if B is FALSE and logically should not stop, then the system continues to trace the source of A upwards. Finally, paths that can logically explain the abnormal node state are retained, while bypass branches belonging to normal control logic or unrelated to the current anomaly are eliminated.

[0046] Subsequently, all the key causal paths (including nodes and edges on the paths) selected above are merged and extracted from the dynamically activated subgraph to form an independent fault subgraph. This subgraph is usually presented as a tree or directed acyclic graph (DAG) structure, with the root node typically representing the root cause of the fault (such as a damaged sensor or an incorrect logical branch), and the leaf nodes representing the initially detected anomalies. Finally, the fault subgraph is output in the form of a data structure (such as an adjacency matrix or a JSON object) for front-end interface rendering and display to operations and maintenance personnel.

[0047] In summary, the knowledge graph-based wind turbine PLC fault location method according to embodiments of the present invention is explained. By parsing the wind turbine physical design document and the PLC control program source code, it effectively bridges the heterogeneous semantic gap. Furthermore, it performs real-time dynamic event instantiation and graph embedding on the real-time operating data stream, generating dynamic node embedding vector sequences and dynamic activation subgraphs. This enables the knowledge graph to dynamically capture the wind turbine's operating status, overcoming the limitations of static topology. Subsequently, by introducing a logical-physical consistency propagation verification mechanism to identify abnormal nodes, and performing reverse causal path search and fault subgraph extraction based on the dynamic activation subgraph, it accurately and interpretably locates the root cause of the fault, effectively mitigating the information masking and false alarm problems caused by cascading effects. Thus, it enables more accurate, faster, and smarter fault location for wind turbine PLCs, significantly improving the operational reliability, maintenance efficiency, and overall economic benefits of wind turbine units.

[0048] Furthermore, a knowledge graph-based fault location system for wind turbine PLCs is also provided.

[0049] Figure 3 This is a block diagram of a knowledge graph-based PLC fault location system for wind turbines according to an embodiment of the present invention. Figure 3As shown, the wind turbine PLC fault location system 300 based on a knowledge graph according to an embodiment of the present invention includes: an information acquisition module 310, used to acquire a set of wind turbine physical design documents and PLC control program source code; a knowledge graph generation module 320, used to generate a two-layer static global knowledge graph based on the set of wind turbine physical design documents and PLC control program source code; a real-time dynamic event instantiation and graph embedding module 330, used to perform real-time dynamic event instantiation and graph embedding on the acquired real-time wind turbine operation data stream to obtain a dynamic node embedding vector sequence and a dynamic activation subgraph; a logical-physical consistency propagation verification module 340, used to perform logical-physical consistency propagation verification on the dynamic node embedding vector sequence to obtain an abnormal node set; and a fault subgraph extraction module 350, used to perform reverse causal path search and fault subgraph extraction on the abnormal node set based on the dynamic activation subgraph to obtain a fault subgraph.

[0050] Furthermore, the knowledge graph generation module 320 is specifically used for: constructing a physical layer knowledge subgraph based on the wind turbine physical design document set; constructing a logical layer knowledge subgraph based on the PLC control program source code; and performing cross-domain semantic fusion on the physical layer knowledge subgraph and the logical layer knowledge subgraph to obtain a two-layer static global knowledge graph.

[0051] Furthermore, the knowledge graph generation module 320 is specifically used for: aligning entity anchor points on the physical layer knowledge subgraph and the logical layer knowledge subgraph to obtain a list of cross-domain entity pairs that have passed verification; instantiating cross-layer semantic edges on the list of cross-domain entity pairs that have passed verification to obtain a set of cross-layer connection edges; and aggregating global heterogeneous graph data on the physical layer knowledge subgraph, the logical layer knowledge subgraph, and the set of cross-layer connection edges to obtain a two-layer static global knowledge graph.

[0052] As described above, the knowledge graph-based wind turbine PLC fault location system 300 according to embodiments of the present invention can be implemented in various wireless terminals, such as servers with knowledge graph-based wind turbine PLC fault location algorithms. In one possible implementation, the knowledge graph-based wind turbine PLC fault location system 300 according to embodiments of the present invention can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the knowledge graph-based wind turbine PLC fault location system 300 can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the knowledge graph-based wind turbine PLC fault location system 300 can also be one of many hardware modules of the wireless terminal.

[0053] Alternatively, in another example, the knowledge graph-based wind turbine PLC fault location system 300 and the wireless terminal can also be separate devices, and the knowledge graph-based wind turbine PLC fault location system 300 can be connected to the wireless terminal via wired and / or wireless networks, and transmit interactive information in accordance with an agreed data format.

[0054] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A knowledge graph-based fan PLC fault positioning method, characterized in that, include: Obtain the wind turbine physical design document set and PLC control program source code; A two-layer static global knowledge graph is generated based on the wind turbine physical design document set and PLC control program source code. The acquired real-time wind turbine operation data stream is instantiated with real-time dynamic events and embedded into a graph to obtain a dynamic node embedding vector sequence and a dynamic activation subgraph; Logical-physical consistency propagation verification is performed on the dynamic node embedding vector sequence to obtain the set of abnormal nodes; Based on the dynamic activation subgraph, a reverse causal path search and fault subgraph extraction are performed on the abnormal node set to obtain the fault subgraph.

2. The knowledge graph-based fan PLC fault positioning method according to claim 1, characterized in that, Based on the wind turbine physical design document set and PLC control program source code, a two-layer static global knowledge graph is generated, including: Based on the wind turbine physical design document set, a physical layer knowledge subgraph is constructed; Based on the PLC control program source code, construct a logic layer knowledge subgraph; Cross-domain semantic fusion is performed on the physical layer knowledge subgraph and the logical layer knowledge subgraph to obtain a two-layer static global knowledge graph.

3. The knowledge graph-based fan PLC fault positioning method according to claim 2, characterized in that, Cross-domain semantic fusion is performed on the physical layer knowledge subgraph and the logical layer knowledge subgraph to obtain a two-layer static global knowledge graph, including: Align entity anchor points in the physical layer knowledge subgraph and the logical layer knowledge subgraph to obtain a list of cross-domain entity pairs that have passed verification. The cross-domain entity pair list that passes the verification is instantiated with cross-layer semantic edges to obtain a set of cross-layer connection edges; Global heterogeneous graph data aggregation is performed on the physical layer knowledge subgraph, the logical layer knowledge subgraph, and the cross-layer connection edge set to obtain a two-layer static global knowledge graph.

4. The knowledge graph-based fan PLC fault positioning method according to claim 1, characterized in that, The acquired real-time wind turbine operation data stream is subjected to real-time dynamic event instantiation and graph embedding to obtain a dynamic node embedding vector sequence and a dynamic activation subgraph, including: Edge detection and event tuple generation are performed on the real-time operation data stream of the wind turbine to obtain a list of event tuples; Based on the event tuple list, a dynamic activation subgraph is obtained by activating the two-layer static global knowledge graph using event-driven dynamic logical subgraphs. Spatiotemporal feature fusion and node embedding calculation are performed on the dynamic activation subgraph and the global node embedding vector matrix of the previous time step to obtain the dynamic node embedding vector sequence.

5. The knowledge graph-based fan PLC fault positioning method according to claim 1, characterized in that, Based on a list of event tuples, an event-driven dynamic logical subgraph activation is performed on the two-layer static global knowledge graph to obtain a dynamically activated subgraph, including: The list of event tuples is traversed and indexed to obtain the set of active seed nodes; In a two-layer static global knowledge graph, a multi-hop logical propagation breadth search is performed on each activated seed node in the set of activated seed nodes to obtain the set of effective nodes and the set of effective edges. The dynamic activation subgraph is obtained by instantiating and reconstructing the dynamic attention subgraph on the set of valid nodes and the set of valid edges.

6. The knowledge graph-based PLC fault location method for wind turbines according to claim 1, characterized in that, Logical-physical consistency propagation verification is performed on the dynamic node embedding vector sequence to obtain the set of abnormal nodes, including: Extract the first dynamic node embedding vector from the dynamic node embedding vector sequence; Extract the environmental state node embedding set, logical side expectation vector, and physical side actual feedback node embedding vector associated with the first dynamic node embedding vector from the two-layer static global knowledge graph; The first working condition context vector is obtained by synthesizing the first dynamic node embedding vector and the environmental state node embedding set into a dynamic working condition context vector. Generate the first tolerance matrix based on the first working condition context vector; Based on the first tolerance matrix, the logical side expectation vector, and the physical side actual feedback node embedding vector, the violation score of the first dynamic node embedding vector is determined. Based on the comparison between the violation score and the preset threshold, it is determined whether to include the node corresponding to the first dynamic node embedding vector into the abnormal node set.

7. The knowledge graph-based PLC fault location method for wind turbines according to claim 6, characterized in that, The first working condition context vector is obtained by synthesizing the first dynamic node embedding vector and the set of environment state node embeddings into a dynamic working condition context vector, including: synthesizing the dynamic working condition context vector using the following formula, wherein the formula is: in, Represents the first working condition context vector; Embed a vector for the first dynamic node; Embed a set of associated environment state nodes; and These are the projection weight matrices for the keys and values, respectively; This is the scaling factor; Presentation layer normalization operation.

8. A knowledge graph-based PLC fault location system for wind turbines, characterized in that, include: The information acquisition module is used to acquire the wind turbine physical design document set and PLC control program source code; The knowledge graph generation module is used to generate a two-layer static global knowledge graph based on the wind turbine physical design document set and PLC control program source code. The real-time dynamic event instantiation and graph embedding module is used to perform real-time dynamic event instantiation and graph embedding on the acquired real-time wind turbine operation data stream to obtain dynamic node embedding vector sequences and dynamic activation subgraphs. The logical-physical consistency propagation verification module is used to perform logical-physical consistency propagation verification on the dynamic node embedded vector sequence to obtain the set of abnormal nodes. The fault subgraph extraction module is used to perform reverse causal path search and fault subgraph extraction on the set of abnormal nodes based on the dynamic activation subgraph to obtain the fault subgraph.

9. The knowledge graph-based wind turbine PLC fault location system according to claim 8, characterized in that, The knowledge graph generation module is specifically used for: Based on the wind turbine physical design document set, a physical layer knowledge subgraph is constructed; Based on the PLC control program source code, construct a logic layer knowledge subgraph; Cross-domain semantic fusion is performed on the physical layer knowledge subgraph and the logical layer knowledge subgraph to obtain a two-layer static global knowledge graph.

10. The knowledge graph-based PLC fault location method for wind turbines according to claim 9, characterized in that, The knowledge graph generation module is specifically used for: Align entity anchor points in the physical layer knowledge subgraph and the logical layer knowledge subgraph to obtain a list of cross-domain entity pairs that have passed verification. The cross-domain entity pair list that passes the verification is instantiated with cross-layer semantic edges to obtain a set of cross-layer connection edges; Global heterogeneous graph data aggregation is performed on the physical layer knowledge subgraph, the logical layer knowledge subgraph, and the cross-layer connection edge set to obtain a two-layer static global knowledge graph.

Images